Python: Ensure that we have the GIL in a callback

Python <3.7 required a call to PyEval_InitThreads() before acquiring the
GIL once per thread. Callbacks are potentially issued from a different
thread and therefore we need to call this function in callback routines.
Unfortunately, calling PyEval_InitThreads() is too late in Cython, as
other operations needing the GIL already took place. As workaround, we
cimport cython.parallel, which does the same thing in a slightly
different way (see
cython/cython#2205 (comment)).

Starting with Python 3.7, none of this is required any more, as
PyEval_InitThreads() is always called.
https://vstinner.github.io/python37-gil-change.html has some background
on it.

Fixes #37