BufferedGzipFile in Python3.5

[alvations]

On the CI tests for Python 3.5, BufferedGzipFile test is breaking at https://github.com/nltk/nltk/blob/develop/nltk/test/data.doctest#L327

This is because of the change in how buffer is read when code was modified from Python3.4 -> 3.5:

• Python3.4: https://github.com/python/cpython/blob/3.4/Lib/gzip.py#L189
• Python3.5: https://github.com/python/cpython/blob/3.5/Lib/gzip.py#L174

There's a new GzipFile._buffer variable in Python3.5 that clashes with nltk.data.BufferedGzipFile._buffer, so at https://github.com/nltk/nltk/blob/develop/nltk/data.py#L372 ,

• When initializing a nltk.data.BufferedGzipFile, it uses the gzip.GzipFile.__init__ that sets the ._buffer variable to the file object
• And when nltk.data.BufferedGzipFile sets self._buffer = BytesIO(), the super-class' buffer was overwritten and unloaded the file object within the self._buffer and it reads an empty file causing test.read() in the doctest to be None.

Since by default Python3.5 uses a buffer reader, it should have the same functionality as nltk.data.BufferedGzipFile. Assuming that the core Python team don't backport this functionality, we should consider deprecating this when we start dropping python2.7 support.

Meanwhile I think the easiest fix is to refactor BufferedGzipFile._buffer to do something like BufferedGzipFile._nltk_buffer. And that should resolve Python3.4 -> 3.5 related issues for now =)