Ensure fusion gating is identity when audio disabled

[Robbyswimmer]

Summary

• update the cross-attention block to apply residual updates without forcing a layer-norm so that the base hidden state is unchanged when no audio signal is present
• gate the residual update produced by the fusion adapter, guaranteeing that gate=0 cleanly skips audio fusion while still allowing gradients to flow when audio is active

Testing

• pytest -q

---

https://chatgpt.com/codex/tasks/task_e_68d31a3f3598832a955cac98966a179f