_Tree.clear() is called by _Tree.__setstate__() to reset both .data and
._firstbucket before actually proceeding with setting internal state
with data loaded from database.

However since 0718a45e (Use a Python __reduce__ method to make pickles
match C.) ._data reset became conditional, first checking that ._data is
not already empty.

The .__setstate__() is often called with object being a ghost.  However
ZODB was never promising to keep attributes alive on ghost objects and
in fact attributes kept in __dict__ was always released, but attributes
in slotted variables, probably due to overlook, started to be released
on ghostification only recently:

    https://github.com/zopefoundation/persistent/commit/fe2219f4
    https://github.com/zopefoundation/persistent/pull/44

and that change causes many BTree tests to fail in pure-python part, e.g.:

    ERROR: testGhostUnghost (BTrees.tests.test_IIBTree.IIBTreeTest)
    ----------------------------------------------------------------------
    Traceback (most recent call last):
      File "BTrees/tests/common.py", line 122, in testGhostUnghost
        self.assertEqual(list(root2[i].items()) , list(t.items()))
      File "/home/kirr/src/wendelin/z/persistent/persistent/persistence.py", line 266, in __getattribute__
        oga(self, '_p_activate')()
      File "/home/kirr/src/wendelin/z/persistent/persistent/persistence.py", line 386, in _p_activate
        jar.setstate(self)
      File "/home/kirr/src/wendelin/z/ZODB/src/ZODB/Connection.py", line 771, in setstate
        self._reader.setGhostState(obj, p)
      File "/home/kirr/src/wendelin/z/ZODB/src/ZODB/serialize.py", line 623, in setGhostState
        obj.__setstate__(state)
      File "BTrees/_base.py", line 1069, in __setstate__
        self.clear()
      File "BTrees/_base.py", line 818, in clear
        if self._data:
      File "/home/kirr/src/wendelin/z/persistent/persistent/persistence.py", line 268, in __getattribute__
        return oga(self, name)
    AttributeError: _data

Via unconditionally resetting ._data in clear we can make the problem go away.

After this patch with PURE_PYTHON=1 BTree testsuite is passing fully again.

Point the doc link to the rendered site.

Mostly for PyPI benefit.

[skip ci]

Add requirements file for RTD.

More consistency between Python and C impls.

Fixes #42.

See CHANGES.rst for details on what changes. Additional test cases cover
the changed behaviour.

build BTrees for 3.4

The wineggbuilder does that already.  Eventually, we will drop building them altogether, and when we do, it will be easier to just decommission that machine.

attempt to fix pypi password

install zodb to make sure those dependent tests are run

Use unittest.skip to skip tests

This way skipped tests will be visible in the test runner output.

Add Travis and AppVeyor build badges

(Try to) fix appveyor build

Replace with 'PyLong_AsLongLongOverflow'.

Toward a fix for #32.

See: https://github.com/zopefoundation/BTrees/issues/32.

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

Introduced in PR #20, they get tested via the 'py27-pure' tox environment.

build os x wheels and deploy to pypi