add GeopIP integration test and use cpythons do_mkvalue and fix some problems of test_builtin.py

Unfortunately there are still a few hiding which means we can't yet run test_builtin.py right now

Tuple and list calling order fixing.

add reload() and a integration test for simplejson

in tuple operation, if the op type check failed, return NotImplemented instead throw exception, let custom type has chance to call their own function

throw exception in list multiply only if the second op is indexable, otherwise return NotImplemented

create a unidecode integration test and enable test_imp

this tests the import changes from #1105 and did not work before this change.

Call PyType_Modified() in add_operators()

Sometimes we started filling in the method-cache before calling add_operators.
add_operators directly modifies the class objects, so it needs to signal that
the cache is now invalid.

I think CPython avoids needing this since they never call add_operators
after doing any type lookups, but we are a bit more flexible with our initialization.

fix import

I could not find a reliable way to reproduce it and don't fully understand the problem :-(
But it looked like we were sometimes calling into the generic setattr instead of the classobj setattr when doing a delete.

this fixes some smaller differences we had and adds some previously unsupported stuff to the imp module
(we could not load C extensions and were too pedantic about some args before, ...)
In addition this gives us:
 - import lock
 - we load now a lot of the builtin modules lazely
 - support for printing out which files get loaded in verbose mode
 - we can now delete a builtin module from sys.modules and import it again (some cpython tests use this)

It also made a classobj.__delattr__ problem appear

Add a ReturnConvention for getattrs that can C++ throw

Otherwise callers have no way of knowing if an exception might
occur in the rewritten code, and they'd have to conservatively
not rewrite those cases (such as 3-arg getattr).

Well, previously getattr() wasn't doing that, so it would rewrite
itself correctly for a descriptor that is not throwing now, but would
start throwing later (those exceptions would get propagated instead
of being caught by getattr).

Directly use cpythons errors.c file

switch to cpython traceback implementation

The performance is the same but it removes quite a lot of unnecessary code and makes hopefully maintaining our codebase a little easier.

Get libsass working by fixing c++ EH handling

socket fixes

add the _locale and cPickle modules

the module is much faster than the python implementation
But I had to change the code a little bit because of our GC and more importantly because some of our types have different names etc...
I also noticed that we failed to to set capifunc.__module__ in a lot of cases which made pickle error

The static variables should be fine because this is compiled as shared object which we the GC will automatically scan.

Make the pycrypto integration test run the testsuite

fix misc problems (del nonexisting attr, bool ops, super: support nonexisting attr,...)

The socket implementation used to only call close on a socket when all references where gone.
This works fine for cpython which uses reference counting but for pyston (and pypy) this is a problem.
This works around the problem in a similar way as pypy does.
It also uses the same method names so that 3th party libs which already support pypy will automatically work with pyston too.
(e.g. urllib3)