Add more rigerous non-slow grad accum tests

[muellerzr]

What does this PR do?

Should be merged after #35651 (tests will fail until I rebase).

This PR tweaks the Trainer grad_accum tests to:

• Use a much smaller CLM (nickypro/tinyllama-15M)
• Use much smaller data + chunk of the dataset (max_length=16 + first 40 items)

This takes us on CPU from ~76s per test to ~6s per test so we can run it as part of our PR CI and squash grad accum issues before they get merged

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 @ArthurZucker

[hiyouga]

To check the GA loss for the condition that loss_kwargs was not enabled, could we simply hack the trainer by setting trainer.model_accepts_loss_kwargs=False before trainer.train()?

[muellerzr]

@hiyouga if you notice, diff_broken does this already in both tests.

[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.

[hiyouga]

@muellerzr Yhh I get it. The existing test cases only ensure the "diff broken" is very off, but could we have a test case to ensure the models do not accept loss_kwargs (like Qwen2-VL) to be correctly trained when GA is enabled?

[techkang]

When testing compute_loss_func, I think it's better to use a model that doesn't accept loss kwargs. Maybe TinyStories-33M is ok?"