feat(linked): add vLLM get_cutlass_moe_mm_data provider

[voltjia]

Summary

• Add a linked vLLM _C::get_cutlass_moe_mm_data implementation at NVIDIA slot 16.
• Retain native slot 0 as the complete default implementation.
• Reject provider-unsupported is_gated=false calls and pre-sm90 devices before invoking the provider.

Motivation

The installed vLLM DSO exposes this multi-output CUTLASS MoE metadata operator, but the tested provider build supports only gated MoE and compiles this specific kernel for sm90 or newer. A separate linked slot lets callers opt into the provider subset without weakening or deleting the native implementation.

N/A - no linked issue.

Type of Change

• feat - new feature / new operator / new platform
• fix - bug fix
• perf - performance improvement
• refactor - code restructuring without behavior change
• test - adding or fixing tests only
• docs - documentation only
• build / ci - build system or CI configuration
• chore - tooling, formatting, or other non-code changes
• Breaking change

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

python3 -m pytest tests -m smoke --devices nvidia -q
77 passed, 20 skipped, 7801 deselected
2 unrelated existing GEMM tolerance cases failed in the vLLM 0.10/CUDA 12.8 image.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Release operator-pruned build passed on A100	Focused suite: 15 passed, 10 skipped; 6 legacy descriptor-assert cases deselected because they require assertions-enabled native builds. Slot-16 unsupported-parameter and unsupported-architecture checks passed in Release. Successful provider execution requires sm90+ hardware.
Iluvatar	No	N/A	N/A
MetaX	No	N/A	N/A
Cambricon	No	N/A	N/A
Moore	No	N/A	N/A
Ascend	No	N/A	N/A

Benchmark / Performance Impact

N/A - additive provider slot; no performance claim.

Notes for Reviewers

• Default dispatch remains native slot 0. Callers must explicitly select implementation_index=16 to use this provider.
• There is no automatic parameter-aware fallback. Unsupported linked calls fail with Release-safe C10 checks before tensor conversion or provider launch.
• The exact 10-argument provider schema mutates five output tensors and accepts optional blockscale_offsets; the adapter only reorders them to the InfiniOps output-tail convention.
• The provider DSO contains other sm80 kernels, but its moe_data.cu target excludes sm80. The A100 restriction is specific to this operator build, not the linked mechanism or the whole DSO.