Remove unconditional train_batch_size assignment (#43770)
* Remove unconditional train_batch_size assignment
The train_batch_size should only be saved to TrainerState when
auto_find_batch_size is enabled (which is already handled in the
auto_find_batch_size block at line 2251). The unconditional assignment
caused issues when resuming from checkpoint with different batch
configurations.
Fixes #43708
* Add test for train_batch_size not saved without auto_find_batch_size
* Only restore train_batch_size from checkpoint when auto_find_batch_size is enabled
Fixes #43708
When resuming from a checkpoint, the trainer was unconditionally restoring
the saved train_batch_size, overwriting the user's current batch size
configuration. This caused incorrect max_steps calculation when users
wanted to resume training with a different batch size.
Now the checkpoint's train_batch_size is only restored when
auto_find_batch_size=True, as that feature specifically needs to resume
with the automatically-found batch size. Otherwise, the user's current
args batch size is used.
Added test to verify users can change batch size when resuming.
