Fix/best model checkpoint fix

[seanswyi]

What does this PR do?

Fixes #35609 (TL;DR best_model_checkpoint is being set when the checkpoint may not even exist)

• Add a new attribute to TrainerState called best_global_step that keeps track of the step where we performed evaluation and achieved a new best metric.
• Moved setting best_model_checkpoint from _determine_best_metric to _save_checkpoint.
• Inside _save_checkpoint, check if the best_model_checkpoint actually exists, and set it if so.

This was a bit trickier than I thought it would be, since the Trainer's saving and evaluation logic are not tightly coupled as would be in the save_strategy == "best" scenario.

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. Trainer sets state.best_model_checkpoint even when it doesn't save there; leads to training crash #35609
• 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?

@muellerzr
@SunMarc
@tomaarsen

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.