Now we exec _kwant_version.py into a namespace and get
the version from that.

Specify only Python3 support in trove classifiers.
https://www.python.org/dev/peps/pep-0301/#distutils-trove-classification

Python 3 no longer allows implicit relative imports:
https://docs.python.org/release/3.0.1/whatsnew/3.0.html#removed-syntax
Instead of importing the version file we just parse it. By keeping
the extension ".py" we make it so that it will be distributed along
with the rest of the package.

Running 2to3 removed a line in the graphene.py tutorial script, which
broke the diff file. We manually update the diff file to fix this.

Under Python 3, `scipy.sparse.coo_matrix` fails when passed memoryviews.
We coerce the memoryviews to numpy arrays before passing them to the
`coo_matrix`. Even though this is kind of a hack it shouldn't be very
inefficient as no data is actually copied. See
https://github.com/scipy/scipy/issues/5123 for details.

Unbound methods are not longer a thing in Python 3 [1].  Instead,
functions use the descriptor protocol [2] to properly bind the `self`
argument if they are being called as a method. Cython functions do not
have a `__get__` attribute and so don't satisfy the descriptor
protocol. They therefore can't be used as methods. As a workaround we
create a wrapper class with the proper __get__ method. The advantage of
this over creating a Python wrapper is that the function signature is
preserved.

[1]: https://docs.python.org/3.0/whatsnew/3.0.html
[2]: https://docs.python.org/3/howto/descriptor.html

Also, replace redundant check against Callable abc
with builtin `callable` function.

The higher-level wrappers like `decomp_lu` take
strings, and we convert them explicitly inside.

The pip tool messes up the timestamps of files, so with our previous strategy
of enabling cython by default the build fails if Cython is not available.

Also, turn the "warning" into a "caution" message that is less frightening and
send it to stdout only, so that it does not appear with pip.