feat: add Nemotron 3.5 S5000 fused-MoE config

[yeahdongcn]

Summary

Make the torchada Triton fused-MoE tuning path safe for Nemotron 3.5 Lightning BF16 (E=128,N=1856, non-gated relu2_no_mul) on S5000. The branch removes the unvalidated auto-loaded JSON candidate, prevents stale routing metadata from crossing graph replay, rejects the non-graph-safe Python alignment fallback during capture, and fixes tuner input reuse and packed non-gated shapes.

Validation

• tests/test_moe_pipeline.py tests/test_nemotron_moe_config.py: 14 passed.
• Pre-commit hooks: passed.
• The prior standalone graph result used a routing cache and is not retained as correctness evidence. A new GPU graph test must cover route mutation and compare each replay with an eager reference before adding a config.
• An independent vLLM-MUSA target-only diagnostic on S5000 (driver 5.2.0, torch 2.11.0.post1+musa5.2.0, Triton 3.2.0) loaded the 61.31 GiB checkpoint and produced deterministic semantic output with graph mode NONE; this does not qualify the torchada kernel or a full graph service.

Pending gate

No Nemotron autotune JSON is currently shipped from this branch. Before merging, run the tuner on S5000 with native alignment, dynamic route graph replay, finite/reference output checks, and measured per-M bucket results. Then integrate only the candidates that pass those gates and re-run vLLM-MUSA with PCG and full graph enabled.