feat(nvidia): link vLLM topk_softmax provider

[voltjia]

Summary

• Add a linked vLLM _moe_C::topk_softmax implementation at NVIDIA slot 16.
• Retain native slot 0 and the existing public API as the complete default path.
• Reject is_padding at the linked boundary because the tested provider schema does not expose it.

Motivation

Vendor vLLM 0.20.2 exposes a six-argument topk_softmax C10 operator. Registering it as a separate implementation proves the _moe_C provider path while preserving InfiniOps behavior for the parameter that this provider version lacks.

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
79 passed, 20 skipped, 7826 deselected

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Release operator-pruned build and smoke passed on A100	Focused suite: 83 passed. Provider matrix covered 3 input dtypes, 3 index dtypes, optional bias, and both renormalization modes.
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. Slot 16 is selected explicitly and has no automatic fallback.
• is_padding is checked before tensor conversion and rejected with TORCH_CHECK, so Release and Debug behavior match.
• The YAML records the exact six-argument schema registered by the tested Verl vLLM 0.20.2 DSO and requires the CUDA dispatch key.