Minor test addition for sdpa producing NaNs for pad tokens

[DuyguA]

The issue was fixed from PyTorch side, still I added a quick test to confirm the issue is indeed resolved.

Fixes #31035

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• [ x] Did you read the contributor guideline,
  Pull Request section?
• [ x] 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.
• [ x] Did you write any new necessary tests?

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: bert

[Rocketknight1]

cc @Cyrilvallez for attention

[Cyrilvallez]

Hey! Thanks for opening the PR. However, the issue you are referring to is very old and well-known, and we've been working around it for quite some time already. Moreover, our tests run on latest pytorch, so it does not actually test anything.
Closing the PR for now, feel free to reopen if you think we missed something! 🤗

[DuyguA]

Hey! Thanks for opening the PR. However, the issue you are referring to is very old and well-known, and we've been working around it for quite some time already. Moreover, our tests run on latest pytorch, so it does not actually test anything. Closing the PR for now, feel free to reopen if you think we missed something! 🤗

Sure, no worries @Cyrilvallez 🙂 One tiny question, for some reason flash attention is not supported for BERT and T5, I wonder PRs per model would be of interest?

[Cyrilvallez]

Hey! Bert got refactored a few days ago in #38301 to support it. Still not the case for T5 though if you want to try it out!