transformers

[Qwen3.5] Fix GDN linear attention multi-token cached forward

#45513

Merged