fix(opt): preserve attention_mask=None in create_causal_mask for SDPA torch.compile

[ArjunPakhan]

Fixes #48924

Summary

When attention_mask is None, OPT was instantiating an all-ones attention_mask tensor for positional embeddings and overwriting the variable before calling create_causal_mask. This caused create_causal_mask to receive a concrete 2D tensor mask instead of None, preventing SDPA from relying on is_causal=True and triggering Triton kernel fallbacks under torch.compile.

Solution

Decoupled positional embedding mask calculation into pos_attention_mask. This ensures attention_mask remains None when passed into create_causal_mask, preserving the fast-path SDPA tracing logic.

Testing

Ran local OPT unit test suite:
pytest tests/models/opt/test_modeling_opt.py -> 139 passed.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: qwen3_omni_moe

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 35385952536:2
Result: success | Jobs: 16 | Tests: 190,609 | Failures: 0 | Duration: 17h 8m