feat(linked): add vLLM moe_wna16_gemm provider

[voltjia]

Summary

• Add a linked vLLM _moe_C::moe_wna16_gemm implementation at NVIDIA slot 16.
• Retain native slot 0 and the existing public API as the default implementation.
• Adapt the provider's output-second, tensor-returning schema to InfiniOps output-last, void-return execution semantics.

Motivation

The vendor vLLM DSO registers moe_wna16_gemm with the C10 Dispatcher. Calling that provider directly avoids maintaining another provider path while preserving InfiniOps tensor ordering and stream control.

#906 is merged and provides the shared vllm/_moe_C.*.so library mapping.

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
79 passed, 19 skipped

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Release operator-pruned build and smoke passed on A100	All slot-16 parameterized cases: 67 passed. Provider-specific output check: 1 passed.
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

• Importing the provider registers this operator with the normal C10 Dispatcher; no JIT-specific call path is used.
• The adapter keeps the caller-provided stream through the shared C10 stream bridge.
• The provider mutates its second argument and returns a tensor. InfiniOps passes the output tensor in the provider position, validates one returned value, and keeps the public output-tail/void contract.
• Default dispatch remains native slot 0; linked slot 16 is selected explicitly and has no automatic fallback.