feat(linked): add vLLM topk_sigmoid provider

[voltjia]

Summary

• Add a linked NVIDIA slot 16 for vLLM _moe_C::topk_sigmoid.
• Preserve the existing native slot 0 and public InfiniOps API.
• Reject provider-unsupported padding and scaling combinations, and short-circuit empty batches before provider dispatch.

Motivation

The installed vLLM provider exposes topk_sigmoid through the C10 Dispatcher.
This PR adds a thin linked adapter while retaining the native implementation for
the full InfiniOps contract.

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

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)
• Hygon (WITH_HYGON)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

Remote environment: A100 with the Verl vLLM 0.20.2 provider image, using an
operator-pruned Release/NDEBUG build with WITH_LINKED=ON.

python -m pytest tests -m smoke --devices nvidia -q
79 passed, 20 skipped

The smoke allowlist is unchanged.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Offline Release operator-pruned build/link/install passed on A100	Linked slot 16 focused matrix: 40 passed in 1.15s; earlier native+linked regression: 46 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 - no performance claim.

Notes for Reviewers

• The linked provider's registered six-argument schema supports bias and
  renormalization, but does not expose InfiniOps' is_padding or
  routed_scaling_factor arguments.
• Slot 16 therefore requires is_padding=None and
  routed_scaling_factor == 1.0; both capability boundaries use
  TORCH_CHECK, including Release builds. Native slot 0 remains the default
  and supports the complete API.
• The provider emits a zero-grid CUDA error for an empty token batch. The
  adapter validates metadata/capabilities and then returns before stream or
  tensor conversion when
  um_tokens == 0`; the focused test verifies a
  subsequent CUDA operation and synchronization remain clean.
• The focused matrix covers fp16/bf16/fp32, int32/uint32/int64 indices, bias
  present/absent, renormalization, a non-default stream, empty input, and both
  unsupported-argument checks.
• Provider-specific cases in tests/test_topk_sigmoid.py explicitly select slot 16.
  Builds without the provider skip only those cases after checking the active
  implementation list; the native cases remain in the same module.