Fix seq2seq collator padding

[vasqu]

What does this PR do?

DataCollatorForSeq2Seq currently only supports the longest and silently the no_padding strategies. This PR adds some checks to ensure the padding strategy is respected if valid (e.g. max_length without a given max length defaults back to longest).

Furthermore, added tests for the collator in the style of the DataCollatorForTokenClassification for all kinds of supported datatypes (pt, tf, np).

Something to note: The seq2seq collator causes side effects on the features' labels that are passed. I don't think it's a big deal but thought I'd clarify why I produce features twice within a single test.

Fixes #30521

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?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@Rocketknight1 @amyeroberts

[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.

[Rocketknight1]

Code changes LGTM, and I appreciate the extensive testing! cc @amyeroberts for core maintainer review.

The point you made about side-effects on the input is interesting, though - I agree that it's unlikely to be critical (because people rarely reuse the pre-collator inputs after calling the collator), but it's still suboptimal. Maybe we could consider having our collators make copies of the inputs instead so they don't overwrite them, though this would definitely be a task for a separate PR.

[amyeroberts]

Thanks for fixing and for all of these tests - beautiful!

[amyeroberts]

The point you made about side-effects on the input is interesting, though - I agree that it's unlikely to be critical (because people rarely reuse the pre-collator inputs after calling the collator), but it's still suboptimal. Maybe we could consider having our collators make copies of the inputs instead so they don't overwrite them, though this would definitely be a task for a separate PR.

@Rocketknight1 @vasqu Very much aligned on this - side-effects should definitely be avoided, and it's a bigger piece of work for a follow-up PR.

[vasqu]

@amyeroberts @Rocketknight1 I'm not sure which collators affect which inputs via side-effects so I'd suggest writing a mutability test which assures this wanted behaviour. Afaik, it currently only affects the labels when using the seq2seq collator but maybe I'll catch some others. Got some time now, will be looking into it in a bit.

[amyeroberts]

@vasqu Awesome ❤️ looking forward to seeing the PR!