Implement to_device() for python bindings, add unit tests

[ozancaglayan]

This will fix the issue #1575

• Expose ctranslate2.Device enum to Python
• Implement to_device() method that wraps the overloaded StorageView::to() variant
• Write unit tests

Caveats:

• Using a CPU StorageView and calling .to_device(Device.cpu) seems to create a copy of the backing store. I added an if () on the C++ side but didn't seem to work. I'm not super familiar with C++ and pybind so maybe there's a way to just reflect back the tensor/storage itself in that case.
• When copying a CUDA Storage to CPU, would the original CUDA storage be garbage collected, cleaned up if it gets out of scope or would that risk leaking CUDA memory?

[ozancaglayan]

I don't know if the failure to build win wheels is related or not. Logs does not seem to provide anything useful.

OK found it: pypa/cibuildwheel#1741

[ozancaglayan]

Hi @minhthuc2502 , would it be possible to review this?

[minhthuc2502]

Could you fix the pipeline please?

[ozancaglayan]

Yes, I'll look at it now

[ozancaglayan]

I just merged master in, I can't see any particular reason other than a timeout in the previous pipeline run. Let's see what happens now.

[ozancaglayan]

ok apparently this was a glitch in the build process rather than an issue in the PR itself.

[minhthuc2502]

LGTM! Thanks for your PR!

[ozancaglayan]

Thanks! Do you have any plans of doing a release soon?

[minhthuc2502]

We will do the release ASAP