verify_done: wait not synchronize

[hnyls2002]

Use event.wait() (stream-level) instead of .synchronize() (CPU block) in maybe_wait_verify_done. Schedule-stream prep ops following the wait are ordered after the forward-stream verify via the stream wait; CPU is no longer blocked. Subsequent .cpu() / .item() calls naturally sync the stream.

---

CI States

Latest PR Test (Base): ⏳ Run #26097243050
Latest PR Test (Extra): ❌ Run #26097242813

[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!

[hnyls2002]

/rerun-test test_dsa_dsv32_tp_mtp.py

(2 tries)

[github-actions]

🚀 8-gpu-h200 (1 test): ❌ View workflow run

cd test/ && python3 registered/dsa_models_e2e/test_dsa_dsv32_tp_mtp.py

[github-actions]

🚀 8-gpu-h200 (1 test): ✅ View workflow run

cd test/ && python3 registered/dsa_models_e2e/test_dsa_dsv32_tp_mtp.py

[hnyls2002]

/tag-and-rerun-ci extra

[hnyls2002]

The test_lora_qwen3_8b_logprob_diff.py failure on this branch is a pre-existing cutlass install issue, not caused by this PR. Verified by dispatching the same test on latest main (which contains the cutlass install fix #25756): https://github.com/sgl-project/sglang/actions/runs/26127841829 — passed. Need to merge latest main into this branch to pick up the fix.