Instances of NEOCluster were not deleted as soon as the only referrers were
weak proxies (at least that's what a quick check with the 'gc' module showed
at the beginning of tearDown). In some cases, __del__ was called while the next
test was logging a message, which led to deadlocks.

Without those proxies, it may be reliable, but only on CPython. See
  http://doc.pypy.org/en/latest/cpython_differences.html#differences-related-to-garbage-collection-strategies

Relying on __del__ to close a cluster was wrong. NEOCluster is now a context
manager that does it explicitly at exit, in addition to automatically stop it.
The NEOCluster.stop method combines the previous stop/__del__/reset methods.

A new 'with_cluster' decorator is also added to avoid excessive indentation
in tests. Unindentation of existing tests will be done later.

This is important when using --loop, otherwise tearDown is slower and slower
at removing packets from the log.

testExternalInvalidation is splitted to minimize reindentation.

It is extended to check that the storage is only notified about the
transactions that existed at the time it asked for them. Otherwise,
Replicator.transactionFinished would be called more than once, and
`self.ttid_set.remove(ttid)` would raise KeyError.

The functional version also contained an annoying 'sleep(10)'.

Leaks in filter_queue caused deadlocks in the following threaded tests that
filter connections.

The removed tests only covered this.

The removed test_answerStoreObject_{1,2} only covered the
'raise NEOStorageError', which is already an assertion.

This fixes the following case when the backup is far behing the upstream DB,
and there are transactions being committed at the same time:

1. replicate partition 0
2. replicate partition 0
3. replicate partition 1
4. replicate partition 0
5. replicate partition 1
6. replicate partition 2
7. replicate partition 0
...
and so on in a quadratic way.

When the upstream activity was too high, the backup could even be stuck looping
on the first partitions.

The issue happens when there were commits while the backup cluster was down.
In this case, the master thinks that these commits are already replicated,
reporting wrong backup_tid to neoctl. It solved by itself once:
- there are new commits triggering replication for all partitions;
- all storage nodes have really replicated.

This also resulted in an inconsistent database when leaving backup mode during
this period.

A backup master crashed with the following traceback after a reconnection:

    Traceback (most recent call last):
      File "neo/master/app.py", line 127, in run
        self._run()
      File "neo/master/app.py", line 147, in _run
        self.playPrimaryRole()
      File "neo/master/app.py", line 348, in playPrimaryRole
        self.backup_app.provideService())
      File "neo/master/backup_app.py", line 123, in provideService
        poll(1)
      File "neo/lib/event.py", line 126, in poll
        to_process.process()
      File "neo/lib/connection.py", line 500, in process
        self._handlers.handle(self, self._queue.pop(0))
      File "neo/lib/connection.py", line 110, in handle
        self._handle(connection, packet)
      File "neo/lib/connection.py", line 125, in _handle
        handler.packetReceived(connection, packet)
      File "neo/lib/handler.py", line 117, in packetReceived
        self.dispatch(*args)
      File "neo/lib/handler.py", line 66, in dispatch
        method(conn, *args, **kw)
      File "neo/master/handlers/backup.py", line 52, in invalidateObjects
        app.invalidatePartitions(tid, partition_set)
      File "neo/master/backup_app.py", line 257, in invalidatePartitions
        self.triggerBackup(node)
      File "neo/master/backup_app.py", line 281, in triggerBackup
        assert cell_list, offset
    AssertionError: 0

Many "unit" tests (!= "threaded" tests) don't do more than checking
implementation details, and increase coverage artificially. As with testEvent
in commit 71e30fb9, most of these tests will
either be removed or rewritten as threaded tests.

The fact that the remaining unit tests actually cover code that other test
don't gives motivation to maintain them. It will be also less code to update
when switching to https://pypi.python.org/pypi/mock

I proceeded as follows:

1. Measure coverage for all tests except unit tests. While checking my work,
   I found that coverage stats for threaded/functional/zodb tests are quite
   unstable, so I restarted from the beginning by doing this measure several
   times and only keeping the intersection of coverage data.

2. Measure coverage individually for each 'unit' tests, and substract the
   each result with the data in 1.

3. The candidates for deletion are those without any code covered.

Tests I didn't delete:

- neo.tests.master.testElectionHandler: I always do minimal changes about
  election, as long as there's no serious review.

- neo.tests.master.testMasterPT.MasterPartitionTableTests.test_13_outdate

- 4 tests in neo.tests.testPT:
  test_01_Cell, test_04_removeCell, test_06_clear, test_08_filled

- neo.tests.storage.testStorage{MySQL,SQLite}

- neo.tests.testUtil.UtilTests.testReadBufferRead

In a way, this commit is actually quite conservative. There are still many
useless tests that only check error paths and for simple tested methods, this
is just duplicating thie tested code.