On ._p_deactivate() and ._p_invalidate(), when an object goes to ghost
state, objects referenced by all its attributes, except related to
persistence machinery, are released, this way freeing memory (if they
were referenced only from going-to-ghost object).

That's the idea - an object in ghost state is simply a stub, which loads
its content on first access (via hooking into get/set attr) while
occupying minimal memory in not-yet-loaded state.

However the above is not completely true right now, as currently on
ghostification only object's .__dict__ is released, while in-slots objects
are retained attached to ghost object staying in RAM:

    ---- 8< ----
    from ZODB import DB
    from persistent import Persistent
    import gc

    db = DB(None)
    jar = db.open()

    class C:
        def __init__(self, v):
            self.v = v
        def __del__(self):
            print 'released (%s)' % self.v

    class P1(Persistent):
        pass

    class P2(Persistent):
        __slots__ = ('aaa')

    p1 = P1()
    jar.add(p1)
    p1.aaa = C(1)

    p2 = P2()
    jar.add(p2)
    p2.aaa = C(2)

    p1._p_invalidate()
    # "released (1)" is printed

    p2._p_invalidate()
    gc.collect()
    # "released (2)" is NOT printed     <--
    ---- 8< ----

So teach ghostify() & friends to release objects in slots to free-up
memory when an object goes to ghost state.

NOTE PyErr_Occurred() added after ghostify() calls because
pickle_slotnames() can raise an error, but we do not want to change
ghostify() prototype for backward compatibility reason - as it is used
in cPersistenceCAPIstruct.

( I hit this bug with wendelin.core which uses proxies to load
  data from DB to virtual memory manager and then deactivate proxy right
  after load has been completed:

  https://lab.nexedi.com/nexedi/wendelin.core/blob/f7803634/bigfile/file_zodb.py#L239
  https://lab.nexedi.com/nexedi/wendelin.core/blob/f7803634/bigfile/file_zodb.py#L295 )

Fix #38 by returning the correct type from TimeStamp_hash.

CI for travis and appveyor

TimeStamp: sub-second precision for Python implementation of timeTime()

The fractional part of return values was always 0.

Add support for Python 3.5.

Don't use/poison global wheel cache when building w/ PURE_PYTHON.

[ci skip].

Drop support for Python 2.6 and 3.2.

- 2.6 is long out-of-maintenance, and a security quagmire.

- 3.2 cannot be tested on Travis.

[ci skip]

[ci skip]

Fixes #27.

Test plan:

- run detox to generate ALL THE WEIRD FILES
- run check-manifest to see if any made it into the sdist

Support for ZODB on PyPy

Remove aliasing of PickleCache's target_size/cache_size; just use cache_size as specified in the interface.

Replace all remaining uses of 'self.__X' in persistence.py with _OGA/OSA for consistency; also avoid some duplicate calls through __getattribute__ which were more obvious in the OGA form.