[TRTLLM-12557][feat] WideEP FT: add AlltoAll watchdog (1a.4)

[chienchunhung]

Summary

• Add an opt-in AlltoAllWatchdog host thread that polls dispatch/combine completion flags in FIFO order and reports timed-out ranks.
• Wire optional EP health/watchdog support into MoeAlltoAll and NVLinkOneSided, including active-rank-mask forwarding from health.
• Add focused watchdog unit coverage for completion, timeout reporting, existing failed-rank filtering, local-rank handling, FIFO clearing, and workspace-offset reads.

Validation

• python -m compileall -q tensorrt_llm/_torch/alltoall_watchdog.py tensorrt_llm/_torch/distributed/moe_alltoall.py tensorrt_llm/_torch/modules/fused_moe/communication/nvlink_one_sided.py tests/unittest/_torch/modules/test_alltoall_watchdog.py
• git diff --check

Not run: pytest, because the available local Python runtimes do not have torch installed.

[chienchunhung]

Superseded by NVIDIA#15524.