In case of MPS device also copy batch to CPU

[hyenal]

What does this PR do?

This PR fixes an issue when using MPS device and composer's classifier. When computing metrics, outputs is copied to the cpu but not batch which creates some downstream error.

What issue(s) does this change relate to?

Fixes #3094

Before submitting

• Have you read the contributor guidelines?
• Is this change a documentation change or typo fix? If so, skip the rest of this checklist.
• Was this change discussed/approved in a GitHub issue first? It is much more likely to be merged if so.
• Did you update any related docs and document your change?
• Did you update any related tests and add any new tests related to your change? (see testing)
• Did you run the tests locally to make sure they pass?
• Did you run pre-commit on your change? (see the pre-commit section of prerequisites)

[mvpatel2000]

Thanks for the PR!

I think there's a minor error -- once fixed looks good to me! Please feel free to rerequest when this is merged

[mvpatel2000]

Hm... labels can be a complex type so directly calling to seems to fail tests

[hyenal]

To ensure compatibility with HuggingFaceModel I had to create a new method to shift labels outside of eval_forward.

I hope this is not a breaking change, it felt a bit more natural to modify the labels outside of this method as well but there may be reasons for doing it there that I am unaware of.

In addition update_metric will be slightly less efficient although the shift label operation should be very cheap