Add support for moving PyTorch Tensors from python to rust via PyO3

[sebastiandero]

All Credit goes to @egordm

@LaurentMazare Please let us know what you need from us to merge this, we urgently need this functionality

Updated version of #457
Merged changes and added macOS support (dylib)

[egordm]

Good job.
I have just tested it as well. Works well.
I have highlighted a few parts I may have accidentally left behind.

[egordm]

This can be removed.

[egordm]

I think this part can be deleted. :P

[sebastiandero]

Thanks for catching this! Hope we can merge this today @LaurentMazare?

[LaurentMazare]

I've pushed a more barebone sets of changes that should hopefully make it possible to write python extensions using tch, see #704. I think it's somewhat similar but more minimalist, the idea is to only provide some low level api for now and not bring in the dependency to PyO3 explicitely and refine this over time (or have the actual Python glue bits live in another crate).
An upside of the changes is that it lifts a recent overhaul of the build script so as to make it easy to ensure that the Python libtorch version is actually being used rather than potentially a different shared library than what the Python side uses.
A very simple example is in examples/python-extension.

[LaurentMazare]

Closing this now, hopefully the main functionalities are now covered in a slightly different way in the newly minted pyo3-tch crate #730 . The sample Python extension can also be found in a separate repo tch-ext.