What does this PR do?

This PR adds separate checkpoint limits for regular checkpoints and best model checkpoints in the Trainer.

Changes:

• Adds save_total_limit_best parameter to TrainingArguments to control the number of best checkpoints kept
• Separates checkpoint management logic so that best model checkpoints and regular interval checkpoints are tracked independently
• Best checkpoints (based on metrics) are no longer deleted when regular checkpoints are rotated

Motivation:

Currently, save_total_limit applies to ALL checkpoints, meaning best models can be deleted to make room for new regular checkpoints. This PR allows users to preserve their best-performing models while limiting regular checkpoint storage.

Fixes #37196

Before submitting

• 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?

@vasqu