[MoonEP] Add single-flight async event chaining

[wirybeaver]

Summary

• run MoonEP dispatch, BF16 weight prefetch, and combine on its communication stream, with GPU-side event waits at each compute consumer boundary
• enforce process-wide single-flight dispatch → prefetch → combine ordering and poison uncertain queued failures
• reject MoonEP with TBO/SBO and cover async, synchronous fallback, ordering, ownership, and validation behavior with CPU tests

Related PRs

• Builds on and targets the MoonEP integration branch from [MoonEP] BF16 PoC integration for Kimi-K3 #33249.
• Establishes stream/event ordering needed by the MXFP4 follow-up in [MoonEP] MXFP4 experts for Kimi K3 on the DeepGEMM runner in symmetric memory #34874; MXFP4 integration remains out of scope here.

Test plan

• PYTHONPATH=python python -m pytest -q test/registered/unit/layers/moe/test_moonep_buffer.py test/registered/unit/server_args/test_server_args.py (158 passed, 17 subtests passed)
• Black, isort, Ruff, git diff --check, and Python compilation
• Multi-GPU NVSwitch BF16 synchronous/async parity validation

---

CI States

Latest PR Test (Base): ❌ Run #32540483710
Latest PR Test (Extra): ❌ Run #32540483551
Latest PR Test (AMD ROCm 10): ❌ Run #32540483763