Fix fsdp2 for gemma models with shared kv states (#45912)
* fix
Co-authored-by: Charly21r <208873065+Charly21r@users.noreply.github.com>
* accelerate checks Mapping, not dict
---------
Co-authored-by: Charly21r <208873065+Charly21r@users.noreply.github.com>