transformers

fix _resize_token_embeddings will set lm head size to 0 when enabled deepspeed zero3

#26024

Merged