feat: add Moore Threads multi-GPU Suite E support

[Keke946]

Summary

• add a content-addressed Moore Threads vLLM-MUSA runner with MTT S4000 multi-GPU tensor-parallel support
• register MUSA devices explicitly with Ray and bypass the NVIDIA/NVML custom all-reduce path for multi-GPU runs
• add an idempotent compatibility patch for vLLM-MUSA 0.4.2 / MCCL 2.7 to skip zero-element tensor payload broadcasts during profiling
• preserve the orchestrating runner ID when Suite E merges resumable per-chip results
• include validated Suite E scaling results for 1x, 2x, and 4x MTT S4000 GPUs

Result

Configuration	Best output throughput	Per-chip throughput	Speedup vs. 1x	Scaling efficiency
1x MTT S4000 48 GB	410.60 tok/s	410.60 tok/s	1.00x	1.000
2x MTT S4000 48 GB	473.16 tok/s	236.58 tok/s	1.15x	0.576
4x MTT S4000 48 GB	532.61 tok/s	133.15 tok/s	1.30x	0.324

Runtime precision is reported as FP16 because the legacy vLLM-MUSA 0.4.x stack maps requested BF16 execution to FP16.

Validation

• clean Suite E run completed for accuracy and 1x/2x/4x offline scenarios
• top-level and per-chip results record moorethreads_vllm_musa_e0266927
• run ID: 559dc817
• python runners/validate_submission.py --dir results/community/mtt_s4000x4_suite_E_moorethreads_vllm_musa_e0266927_559dc817 passes
• the compatibility patch was applied to a clean upstream communication_op.py in a temporary directory and reproduced the tested installed file byte-for-byte

Known limitation

• MTT power telemetry is not yet integrated, so the validator warns that tokens-per-watt cannot be computed.