perf: tune Nemotron MTP MoE decode on S5000

[yeahdongcn]

Status

This draft carries the shape map that produced the historical Nemotron MTP
result, plus a Triton capability guard. A clean vLLM-MUSA integration gate is
still required before merge.

Changes

The PR packages the S5000 Nemotron non-gated MoE autotune map for
E=128,N=1856,device_name=MTT_S5000 with measured runtime and graph buckets:

16, 49, 80, 113, 144, 511, 512, 513

The serving routes are M≈115/120 -> 113 and M=512 -> 512; values above
the measured range use the generic heuristic. The packaged JSON SHA256 is
a0b192a4625e71cf46c52cca3bf5f042d9988da927524d7c9e4d10a0bbabc983.

The descriptor capability check now requires the actual Triton language API
triton.language.make_tensor_descriptor. Triton 3.2 can expose the tools
module without exposing that constructor; it must use the fallback path. Triton
3.6 continues to enable the descriptor path when the API is present.

Validation

• Candidate commits: 668bd68, f4e2ffb.
• Clean wheel builds succeed and package the JSON above.
• Focused config/resolver and descriptor tests: 10 passed on the MUSA image.
• The full torchada suite reached 508 passed, 19 skipped; the only reported
  failure was the project-version check from a source-only staging harness that
  did not include project metadata.
• The historical manual-injection A/B on the original post1/Triton 3.2 setup
  produced exact-output results:

path	TPOT
target-only	19.6128 ms/token
MTP1	11.4567 ms/token
MTP2	8.8937 ms/token
MTP4	7.1127 ms/token
MTP6	6.6091 ms/token

Those numbers came from the eight-key JSON injected into the installed runtime;
they are historical evidence, not yet a clean torchada==0.1.88 claim.

A diagnostic clean-source run confirmed that the candidate JSON was loaded and
that runtime M=112/126 selected key 113. Its warm-cache result was about
9.46 ms/token on a different S5000 host. A fresh post1/Triton-3.2 compile path
also exposed the descriptor compatibility issue now fixed by f4e2ffb. Neither
diagnostic number is accepted as the final performance result.

The PR should remain Draft until the packaged wheel is run with current
vLLM-MUSA source on the same hardware under a clean, interleaved
baseline-vs-candidate gate, with output hashes, selected MoE buckets,
acceptance, and TPOT recorded.