feat: support Nemotron non-gated MoE tuning on Triton 3.6

[yeahdongcn]

This draft adds the torchada-side support needed to tune Nemotron 3.5 Lightning's non-gated MoE under Triton 3.6.

Changes:

• recognize Triton 3.6.0 when selecting fused-MoE configuration directories;
• model Nemotron's relu2_no_mul route as non-gated, with E=128/N=1856 projection and quantized-K shapes;
• keep routed weighting, activation, reduction, and graph-capture metadata consistent with the runtime path;
• add focused configuration and pipeline tests;
• add S5000 Triton 3.6 configs for Nemotron E=128/N=1856 MoE and Mamba SSU shapes.

Validation:

• py_compile, ruff, JSON validation, and git diff --check pass.
• On leased S5000 GPU0 (driver 3.3.5-server), vLLM-MUSA post2/Triton 3.6 explicitly loaded both config files and completed PCG + decode FULL graph capture.
• Same-host target-only E2E improved from B8 53.693 tok/s without configs to 83.012 tok/s with configs (three measured runs, two warmups, fixed TP1/BF16 workload); B1 was 27.542 tok/s.
• MTP2 graph/request smoke completed at B1 24.079 / B8 41.468 tok/s, but repeated B8 output was not deterministic, so this does not qualify MTP2 for production.
• The candidate config files were initially generated on a rejected host and then revalidated by exact config-load and E2E replay on the correct leased S5000. A clean current-host autotuner rerun was started but interrupted when SOL reclaimed the lease; its incomplete output is not used as evidence.