Fix SHM feature finalization under DP attention

[yinghai]

Summary

• Synchronize on the CPU groups that broadcast SHM-backed multimodal work requests before materializing and unlinking those features.
• Preserve the existing TP-group barrier behavior for the non-DP-attention path.
• Add focused unit coverage for DP-attention, non-DP-attention, and no-SHM finalization behavior.

Testing

• uv run python -m py_compile python/sglang/srt/managers/scheduler_components/request_receiver.py test/registered/unit/managers/test_request_receiver_shm_finalize.py
• uv run python -m pytest test/registered/unit/managers/test_request_receiver_shm_finalize.py (could not run locally: pytest is not installed in this environment)

Source commits

• 5058b632

---

CI States

Latest PR Test (Base): ❌ Run #28308122067
Latest PR Test (Extra): ❌ Run #28308122022

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[yinghai]

/tag-and-rerun-ci