- we can use BINPERSID only if protocol >= 1
- we can use text PERSID only if protocol < 4 and Ref string does not
  contain \n - else encoding have to fail.

Tests for Ref encoding at protocols 3 & 4 will follow after string
encoding is fixed.

- we can use EMPTY_LIST only if protocol >= 1

Also: since we are now using EMPTY_LIST only optionally, it is logical
to switch to

	MARK + ... + LIST

instead of

	EMPTY_LIST (or MARK + LIST @proto=0) + MARK + ... + APPENDS

which is at least 1 byte longer.

For the reference - APPENDS is also from protocol 1, while LIST is from
protocol 0.

- we can use EMPTY_TUPLE only if protocol >= 1
- also: if protocol >= 2 we can now use TUPLE{1,2,3} opcodes.

Corresponding TODO in the code was added in 4fd6be93 (encoder: Adjust it
so that decode(encode(v)) == v)

We can use BINFLOAT opcode only starting from protocol >= 1.
At protocol 0 we must use ASCII FLOAT.

We can use BININT* opcodes only starting from protocol >= 1.
At protocol 0 we must use ASCII INT.

Starting from protocol 2 1-byte NEWTRUE/NEWFALSE opcodes are more
efficient compared to 5-bytes e.g. "I01\n.".

It is not only about efficiency, as protocol 4 _forbids_ use of variable
length ASCII-only opcodes - whose data length is determined by doing
forward scan for '\n'.

Without encodeBool changes and only with the tests added it would fail
this way:

    --- FAIL: TestEncode/True/proto=2 (0.00s)
        ogorek_test.go:383: encode:
            have: "\x80\x02I01\n."
            want: "\x80\x02\x88."
    --- FAIL: TestEncode/True/proto=3 (0.00s)
        ogorek_test.go:383: encode:
            have: "\x80\x03I01\n."
            want: "\x80\x03\x88."
    --- FAIL: TestEncode/True/proto=4 (0.00s)
        ogorek_test.go:383: encode:
            have: "\x80\x04I01\n."
            want: "\x80\x04\x88."
    --- FAIL: TestEncode/False/proto=2 (0.00s)
        ogorek_test.go:383: encode:
            have: "\x80\x02I00\n."
            want: "\x80\x02\x89."
    --- FAIL: TestEncode/False/proto=3 (0.00s)
        ogorek_test.go:383: encode:
            have: "\x80\x03I00\n."
            want: "\x80\x03\x89."
    --- FAIL: TestEncode/False/proto=4 (0.00s)
        ogorek_test.go:383: encode:
            have: "\x80\x04I00\n."
            want: "\x80\x04\x89."

There are many pickle protocol versions - 0 to 4. Python2 for example
understands only versions 0 - 2. However we currently unconditionally
emit opcodes from higher versions, for example STACK_GLOBAL - from
version 4 - when encoding a Class, which leads to inability to decode
pickles generated by ogórek on Python2.

Similarly protocol 0 states that only text opcodes should be used,
however we currently unconditionally emit e.g. BININT (from protocol 1)
when encoding integers.

Changing to always using protocol 0 opcodes would be not good, since many
opcodes for efficiently encoding either integers, booleans, unicode etc
are available only in protocol versions 2 and 4.

For this reason, similarly to Python[1], let's allow users to specify
desired pickle protocol when creating Encoder with config. For backward
compatibility and common sense the protocol version that plain
NewEncoder selects is 2.

This commit adds only above-described user interface and testing
infrastructure for verifying what was the result of encoding an object
at particular protocol version.

For now only a few of pickle test vectors are right wrt what the encoder
should be or currently generates. Thus in the next patches we'll be
step-by-step fixing encoder on this topic.

[1] https://docs.python.org/3/library/pickle.html#pickle.dump

For now all decoding all those pickles is tested to give the same
object, as in e.g.

	"(I1\nI2\ntp0\n.", // MARK + TUPLE + INT

and

	"I1\nI2\n\x86."),  // TUPLE2 + INT

But having a way to specify several pickles to a test case will become
even more handy when later adding precise tests for Encoder - there we
will need to assert that at such and such protocol encoding gives such
and such pickles. And there are 5 protocol versions to test...

Previously there were two separate tables - for decode and encode tests.

The table for encode tests was very small. TestDecode, which was
operating on data from decode table, was also performing checks that
decode(encode(object)) is idempotent - the work which is already too by
TestEncode. However the coverage of input objects from decode table was
not a strict superset of encode table objects.

For the reasons above let's stop this divergence. Let's have a common
table that define tests where for every test case there can be:

	- an "in" object,
	- a pickle, and
	- an "out" object

where

	1. pickle must decode to "out" object, and
	2. encoding "in" object must give some pickle that decodes to "out" object.

	NOTE: In the usual case "in" object == "out" object and they can only
	differ if "in" object contains a Go struct.

This will allow us to cover all existing decode and encode tests logic.

However the coverage of each logic is now higher - for example Encoder
tests are now run on every object from main table, not only for 3 cases
like it was before.

Similarly to TestDecode. No integration with main tests table yet.

We are going to integrate encoding tests with the main tests data table
in several steps.

Only code movement here, no other change.

We are going to modify the main tests table and for every case (e.g.
int(1) add pickles with various encoding that all represent the same
object when decoded. The main TestDecode driver will iterate over all
those input data and hand it over for particular testing to testDecode
worker.

Use t.Run for spawning each case - it is less typing (we do not need to
manually print test.name on errors, etc), and it allows to select which
subtests to run via e.g. `go test -run Decode/int`

Having that long data makes the table clumsy and harder to understand.
By moving such data definition out of the table, we make it a bit easier
to understand.

In the case of "long line", the pickle input and the line itself were
almost duplicating each other, so instead of having two long lines
explicitly pasted, let's have the test input be defined as

+	{"too long line", "V" + longLine + "\n.", longLine},

In the case of Graphite messages, one long Graphite object was also
duplicated in encode_test.go .

Just a cleanup, no semantic change.

It will be handy to use testing.T.Run in the upcoming patches that add
more decode/encode tests. However since testing.T.Run was was added in
Go1.7 the code won't work with Go1.6 .

Since Go1.6 was released in the beginning of 2016 - more than 2.5 years
ago, and upstream Go policy is to support only current stable (currently
Go1.11) and two previous releases (currently Go1.10 and Go1.9), as of
today Go1.6 is both not supported and outdated. So let's drop explicit
support for it in the name of progress.

By above logic we could also drop support for 1.7 and 1.8, but since
there is no pressing particular need to do so I'm not dropping them for
now.

Currently we often use constructs like

	_, err := e.w.Write([]byte{opNone})
	return err

however with added Encoder.emit helper it can become only

	return e.emit(opNone)

which is more clearly readable.

We can do similarly for formatted output:

-	_, err := fmt.Fprintf(e.w, "%c%dL\n", opLong, b)
-	return err
+	return e.emitf("%c%dL\n", opLong, b)

Due to much boilerplate in one place (encodeInt) the return of
fmt.Fprintf was not even checked. We change the code there with

-		_, err = e.w.Write([]byte{opInt})
-		if err != nil {
-			return err
-		}
-		fmt.Fprintf(e.w, "%d\n", i)
+		return e.emitf("%c%d\n", opInt, i)

which is now hopefully more visible for what is going on and is easier
to catch potential mistake by eyes.

A corresponding test for "int64 encoded as string" will be added later.

Protocol 3 adds opcodes to represent Python bytes.
For Protocol 4 we were missing definitions of several of the opcodes,
such as BINUNICODE8, EMPTY_SET etc.

Add definitions for them all.

We already keep "Protocol 2" and "Protocol 4" opcodes separately, and
for a good reason - it must be clear for a person just by looking at
opcodes definition from which protocol version an opcode comes. However
at present Protocol 0 and Protocol 1 opcodes come all intermixed with
each other.

In the upcoming patches we'll be teaching Encoder to take desired pickle
protocol version into account (similarly to how pickle.dumps accepts desired
protocol version in Python), and then the encoder will have to use
different opcodes for different versions: for example if user asks to
produce "protocol 0" pickle stream it will be invalid to use "protocol
1" opcodes - e.g. BININT, EMPTY_TUPLE, etc.

So I went through

	https://github.com/python/cpython/blob/master/Lib/pickletools.py

for each opcode searching its protocol version and separated the opcodes
with proto=1 from protocol 0 opcodes.

It was probably a thinko in d00e99e7 (Add more opcodes so we can read
Graphite's Carbon stream) which changed opcodes from string ("X") to
character ('X') constants and marked the first opcode with byte type,
probably with the idea that following opcodes will have the same type.

Unfortunately it is not so as the following program demonstrates:

	package main

	const (
	        opAAA byte = 'a'
	        opBBB      = 'b'
	)

	func main() {
	        op := opBBB
	        if true {
	                op = opAAA	// <-- line 11
	        }
	        println(op)
	}

	--> ./bc.go:11:6: cannot use opAAA (type byte) as type rune in assignment

Similarly if we try comparing opcodes, it also results in compile error:

	func main() {
	        op := opBBB
	        if op == opAAA {	// <-- line 10
	                panic(0)
	        }
	        println(op)
	}

	--> ./bc.go:10:8: invalid operation: op == opAAA (mismatched types rune and byte)

Since in the following patches it will be handy for encoder to e.g. set default
opcode first, and then change it under some conditions to another opcode,
exactly the same compile error(s) will pop up.

So let's fix all opcode constants to have their type as byte to avoid such
unexpected errors.

For example

	d.push(math.Float64frombits(u))

looks more clear compared to

	d.stack = append(d.stack, math.Float64frombits(u))

and we already use push in many other places.

v is concrete type here (float64) and Go automatically converts it to
interface{} on call to Decoder.push().

While doing a5094338 (encoder: More unexpected EOF handling) I was a bit
disappointed that https://golang.org/cl/37052 was not going to be
accepted and somehow added hack to tests that was doing

	strconv.ErrSyntax -> io.ErrUnexpectedEOF

error conversion if the test name contained "unicode".

Today I was refactoring tests and noticed that it is better we teach loadUnicode
to return io.ErrUnexpectedEOF in the first place, and the next easy way for this
(after fixing strconv.UnquoteChar itself) is to append many "0" to string and
see if strconv.UnquoteChar still returns ErrSyntax.

So do it in our custom unquoteChar wrapper.

Instead of copy-pasting full test driver for each value, make it to be
only one test with many values in table each tested by common driver.

TestZeroLengthData was checking output.BitLen() == 0 which is another
way to test for equality with big.Int(0) according to

	https://golang.org/pkg/math/big/#Int.BitLen

So join it too to the rest of decodeLong tests.

While at decodeLong topic, rename BenchmarkSpeed -> BenchmarkDecodeLong
as this benchmark checks speed of decodeLong only, nothing else.

Add a test that excersizes whether invalid protocol version is caught
and reported. Without 5c420f85 the test fails like this:

	--- FAIL: TestDecodeError (0.00s)
	    ogorek_test.go:312: "\x80\xffI1\n.": no decode error  ; got 1, <nil>

Due to := in

	v, err := d.r.ReadByte()

newly declared err was shadowing outside err and thus even though the
code intent was to treat all protocols != 2 as ErrInvalidPickleVersion,
after leaving the switch err was always =nil and no error was reported.

Fix it by explicitly declaring v and not using := not to shadow err.

We also change the condition for supported protocol version to be in [0, 4] because:

- we already have messages in our testsuite with PROTO opcode and
  version 1, and if we don't adjust the version condition the tests will fail.

- the highest protocol version we support opcodes for is 4.
- even though the documentation for PROTO opcode says version must be >= 2,
  in practice CPython decodes pickles with PROTO and lower version just ok:

	In [18]: pickle.loads("\x80\x02I5\n.")
	Out[18]: 5

	In [19]: pickle.loads("\x80\x01I5\n.")
	Out[19]: 5

	In [20]: pickle.loads("\x80\x00I5\n.")
	Out[20]: 5

  so we don't complain about those lower versions too.

Continuing 84598fe1 (Add support for persistent references) theme let's
develop support for persistent references more: the usual situation (at
least in ZODB context, which is actually the original reason
persistent references exist) is that decode converts a reference of
(type, oid) tuple form into corresponding "ghost" object - an instance
of type, but not yet loaded with database object's data. This way
resulted object tree is created right away with types application
expects (and ZODB/py further cares to automatically load object data
when that ghost objects are accessed).

To mimic this on Go side, with current state, after decoding, one would
need to make a full reflect pass on the resulted decoded objects, find
Refs, and convert them to instances of other types, also caring to
patch pointers in other objects consistently that were pointing to such
Refs. In other words it is some work and it coincides almost 100% with
what Decoder already does by itself.

Thus, not to duplicate that work and conducting parallel with Python
world, let's add ability to configure Decoder and Encoder so that
persistent references could be handled with user-specified application
logic right in the process, and e.g. Decoder would return resulted
object with references converted to corresponding Go types.

To do so we introduce DecoderConfig and EncoderConfig - configurations
to tune decode/encode process. To maintain backward compatibility
NewDecoder and NewEncoder signatures are not adjusted and instead
NewDecoderWithConfig and NewEncoderWithConfig are introduces. We should
be sure we won't need e.g. NewDecoderWithConfigAndXXX in the future,
since from now on we could be adding fields to configuration structs
without breaking backward compatibility.

For decoding there is DecoderConfig.PersistentLoad which mimics
Unpickler.persistent_load in Python:

https://docs.python.org/3/library/pickle.html#pickle.Unpickler.persistent_load

For encoding there is EncoderConfig.PersistentRef which mimics
Pickler.persistent_id in Python:

https://docs.python.org/3/library/pickle.html#pickle.Pickler.persistent_id

( for Persistent{Load,Ref}, following suggestion from @kisielk, the
  choice was made to explicitly pass in/out Ref, instead of raw
  interface{} pid, because that makes the API cleaner. )

Then both Decoder and Encoder are adjusted correspondingly with tests
added.

About tests: I was contemplating to patch-in support for decoder and
encoder configurations, and handling errors, into our main test
TestDecode, but for now decided not to go that way and to put the test
as separate TestPersistentRefs.

By the way, with having configurations in place, we could start to add
other things there - e.g. the protocol version to use for Encoder.
Currently we have several places where we either don't use opcodes from
higher protocol versions (fearing to break compatibility), or instead
always use higher-version opcodes without checking we should be able to
do so by allowed protocol version.

"1.10" is put into quotes because otherwise Travis parses that as 1.1 -
see e.g. https://github.com/travis-ci/travis-ci/issues/9247.

* Add docstrings to types exposed in public API

I was refreshing my knowled of ogórek and added them along the way.

* Add support for persistent references

Python has the mechanism for objects in one pickle to reference objects
in another pickle. This mechanism is primarily used in ZODB object
database for one object to "persistently" reference another object(*).

I needed a way to load and understand objects from a ZODB database in
Go, and hence comes this patch for ogórek that teaches it to understand
such persistent references.

Like it is already with Call and Class we add a dedicated Ref type to
represent persistent references in pickle stream.

(*) in ZODB when one object is changed, there is no need to resave
unchanged object that is referencing it because its reference stays
valid and pointing to changed object: the "persistent ID" in ZODB is
only object ID and which revision of the referenced object it points to
is implicitly follows by current transaction number view of the database.

decoder: Small speedups

readLine is used in many text decoders, so it pays off if we care not to
do allocation for result on every readLine call:

    name      old time/op    new time/op    delta
    Speed-4      378ns ± 0%     379ns ± 1%     ~     (p=0.397 n=5+5)
    Decode-4    70.1µs ± 0%    67.3µs ± 0%   -3.97%  (p=0.008 n=5+5)
    Encode-4    16.6µs ± 0%    16.6µs ± 0%     ~     (p=0.548 n=5+5)

    name      old alloc/op   new alloc/op   delta
    Speed-4       280B ± 0%      280B ± 0%     ~     (all equal)
    Decode-4    32.3kB ± 0%    31.0kB ± 0%   -3.87%  (p=0.008 n=5+5)
    Encode-4    6.54kB ± 0%    6.54kB ± 0%     ~     (all equal)

    name      old allocs/op  new allocs/op  delta
    Speed-4       8.00 ± 0%      8.00 ± 0%     ~     (all equal)
    Decode-4       428 ± 0%       326 ± 0%  -23.83%  (p=0.008 n=5+5)
    Encode-4       297 ± 0%       297 ± 0%     ~     (all equal)

This reduces allocations a bit:

    name      old time/op    new time/op    delta
    Speed-4      383ns ± 4%     378ns ± 0%    ~     (p=0.302 n=5+5)
    Decode-4    72.1µs ± 1%    70.1µs ± 0%  -2.76%  (p=0.008 n=5+5)
    Encode-4    16.5µs ± 0%    16.6µs ± 0%    ~     (p=0.095 n=5+5)

    name      old alloc/op   new alloc/op   delta
    Speed-4       280B ± 0%      280B ± 0%    ~     (all equal)
    Decode-4    35.7kB ± 0%    32.3kB ± 0%  -9.72%  (p=0.008 n=5+5)
    Encode-4    6.54kB ± 0%    6.54kB ± 0%    ~     (all equal)

    name      old allocs/op  new allocs/op  delta
    Speed-4       8.00 ± 0%      8.00 ± 0%    ~     (all equal)
    Decode-4       429 ± 0%       428 ± 0%  -0.23%  (p=0.008 n=5+5)
    Encode-4       297 ± 0%       297 ± 0%    ~     (all equal)

Most of the functions in ogorek.go already use Decoder by pointer, but
NewDecoder() and Decoder.Decode() exceptionally used it by value.

This was probably an oversight because when Decoder struct is used by
value it is copied on every Decode call and also it is not possible to
change and retain state in between calls as all changed happen to copy
and then forgotten.

This also leads to many unnnecessary allocations, i.e. memory allocated
for stack one Decode call is not reused on next Decode call etc.

So use Decoder by pointer only. This leads to the following speedup:

    name      old time/op    new time/op    delta
    Speed-4      377ns ± 0%     383ns ± 4%     ~     (p=0.214 n=5+5)
    Decode-4    80.4µs ± 3%    72.1µs ± 1%  -10.38%  (p=0.008 n=5+5)
    Encode-4    16.5µs ± 1%    16.5µs ± 0%     ~     (p=0.690 n=5+5)

    name      old alloc/op   new alloc/op   delta
    Speed-4       280B ± 0%      280B ± 0%     ~     (all equal)
    Decode-4    44.0kB ± 0%    35.7kB ± 0%  -18.77%  (p=0.008 n=5+5)
    Encode-4    6.54kB ± 0%    6.54kB ± 0%     ~     (all equal)

    name      old allocs/op  new allocs/op  delta
    Speed-4       8.00 ± 0%      8.00 ± 0%     ~     (all equal)
    Decode-4       502 ± 0%       429 ± 0%  -14.54%  (p=0.008 n=5+5)
    Encode-4       297 ± 0%       297 ± 0%     ~     (all equal)

- use existing tests vector as input data
- for Decode we prepare one large pickle stream from testv and decode all pickle from there
- for Encode we prepare one lirge slice with all objects from testv and encode it

On an i7-6600U currently it looks like:

    BenchmarkSpeed-4         5000000               387 ns/op             280 B/op          8 allocs/op
    BenchmarkDecode-4          20000             81892 ns/op           43989 B/op        502 allocs/op
    BenchmarkEncode-4         100000             16711 ns/op            6536 B/op        297 allocs/op

.travis.yml += go1.8

Which is currently latest stable go release. Maybe we should also drop
go1.6 .

Make decode(encode(v)) to be identity + tuple

It is true that tuples and lists are very similar, and when e.g. a
python function accepts tuple as input, and then uses only indexes
access, it can be substituted with list.

However some functions do `isinstance(..., tuple)` etc, and e.g. for a
go program which wants to produce pickle for such python programs there
is currently no way to generate pickle with tuple opcodes.

So to solve this teach ogórek about tuples:

- introduce simple Tuple type which has []interface{} as underlying
- when decoding tuple opcodes produce this Tuple instead of
  []interface{} used previously.
- when encoding encode Tuple with tuple opcodes.

Otherwise Tuple can be seen and behaves like a regular go []interface{}
slice. In fact the only difference is that runtime type of Tuple is
different from runtime type of []interface{} but otherwise both values
are the same and can be casted to each other if/when needed freely.

NOTE opReduce decoder is adjusted to always require tuple, not list,
because with e.g. cPickle:

    "c__main__\nf\n]R." -> TypeError('argument list must be a tuple', ...)
    "c__main__\nf\n)R." -> ok

( the first one uses empty list - "]", and the second one empty tuple - ")" )

So it is ok and compatible to always require args to be tuple for
reduce.