transformers

Fix flex_attention block mask creation when `get_seq_length` returns a tensor

#46802

Merged