Add torch_empty_cache_steps to TrainingArguments

[aliencaocao]

What does this PR do?

Following up on #31530, add torch.cuda.empty_cache() as an optional TrainingArgument to the training loop every N step, such that those who need it to avoid OOMs can use it.

Also added to the reduce vram usage docs and changed the table for mixed precision training a bit - let me know if you prefer this change to be isolated in another PR.

Not sure if a test is needed - we can't really 'verify' it being called besides monitoring vram usage?

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?

@muellerz @amyeroberts

[amyeroberts]

Thanks for adding this!

Overall I think this looks OK, but let's get @muellerzr's opinion

[muellerzr]

Thanks, overall this looks good to me, bar one nit for our docs!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[amyeroberts]

Looks great - thanks for adding!

[muellerzr]

Test failures look unrelated, so we're okay to merge here I think.

[muellerzr]

Well done! Thanks!