transformers

fix(trainer): Correct loss scaling for incomplete gradient accumulation steps

#39659

Merged