Without this, functions won't accept bytearray, memoryview,
or other types that can be exposed as bytes to the C API.
The most resilient way to obtain a bytes-like object from
the C API is using the Py_buffer interface.

This commit converts use of s#/y# to s*/y* and uses
Py_buffer for accessing the underlying bytes array.

I checked how hashlib is implemented in CPython and the
the implementation agrees with its use of the Py_buffer
interface as well as using BufferError in cases of bad
buffer types. Sadly, there's no good way to test for
ndim > 1 without writing our own C-backed Python type.

Differential Revision: https://phab.mercurial-scm.org/D7879