refactor _inner_training_loop to smaller methods (#44041)
* refactor _inner_training_loop to smaller methods
* address PR comments and remove unused args
* move _train_batch_size
* big update
* move DummyScheduler
* style
* make it private
* switch to kwargs_handlers
* style
* fix resuming
* style
* revert loading after
* way better now
* move method
* small mistake
* update
* remove tp_size
* account for grpo
* final touch
---------
Co-authored-by: Marc Sun <marc@huggingface.co>
Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>
