Log the correct learning rate

[SunMarc]

What does this PR do?

The Trainer class steps the learning rate scheduler at the end of each training loop iteration before calling Trainer._maybe_log_save_evaluate(). As a result, the logged learning rate is the learning rate of the next training step, not the one that was just performed. This off-by-one error is usually not much of a problem with monotonic schedulers, but with cycling schedulers such as cosine_with_restarts, it can be the difference between something very close to zero and the max learning rate at cycle boundaries.

Fixes #35942
Similar to #30169

[github-actions]

Hi 👋, thank you for opening this pull request! The pull request is converted to draft by default. The CI will be paused while the PR is in draft mode. When it is ready for review, please click the Ready for review button (at the bottom of the PR page). This will assign reviewers.

[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.

[muellerzr]

Looks good on my end, thanks for giving this a go!

[SunMarc]

This might potentially break trl dpo_online trainer I think @qgallouedec

[qgallouedec]

Thank you! I will check and correct TRL if necessary. But feel free to merge, I can do it once merged.