allow integer device for BatchEncoding

[jethrokuan]

What does this PR do?

Fixes #9244

I'm not fully aware of the details behind the Apex guard, in the method, so maybe this is not the solution.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

tokenizers: @mfuntowicz

[patrickvonplaten]

This looks good to me. torch 1.7 also supports "integer" device assignment. @jethrokuan - do you know if previous PyTorch versions supported this as well?

Also not really sure what is meant by the APEX comment. @sgugger @LysandreJik do you know more here maybe?

[sgugger]

The comment for APEX is because the to method is lso used to convert the dtype of a Tensor, and APEX does not perform type checks before calling the to method (just that the to method exists), which means or BatchEncoding are converted to a dict of half tensors if we don't have this safeguard.

Using an int should work with older versions of PyTorch AFAICT, however it's considered "legacy" behavior so it may be removed in a future version. In the meantime, I have no problem with accepting them on our side.