Suggestion from GvR (one I like): instead of re-using Py_None, add a new
C function that is stored in the tp_hash slot as a sentinel instead of
the Py_None value used in the posted version of the patch. This will
avoid breaking code that just checks for NULL before calling the tp_hash
slot directly, while still allowing typeobject.c to detect that the
object isn't actually hashable despite the presence of a non-NULL value
in the tp_hash slot.

(I'll keep the None at the Python level though, since that matches the
Py3k behaviour and plays nicely with collections.Hashable)