feat(cuda): extend topk_softmax backend coverage

[voltjia]

Summary

• Extend the canonical TopkSoftmax provider to Iluvatar, MetaX, and Moore.
• Move the NVIDIA CUDA kernel into the shared CUDA path while retaining NVIDIA device-guard handling.

Motivation

InfiniCore still uses the deprecated TopkSoftmaxInfinilm compatibility operator on CUDA-compatible backends. Shared canonical provider coverage is required before that adapter can migrate to TopkSoftmax.

Type of Change

• feat - new feature / new operator / new platform
• fix - bug fix
• perf - performance improvement (no behavioral change)
• 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 (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

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

Validated on ssh nvidia in accelerator-dev/nvidia:latest as part of the canonical-provider integration stack:

$ scripts/dev/build.sh nvidia --jobs 16
Build succeeded

$ python -m pytest tests/test_topk_softmax.py --devices nvidia -q
46 passed

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	build passed	46 focused tests passed
Iluvatar	Yes	not run	platform hardware unavailable; platform CI required
MetaX	Yes	not run	platform hardware unavailable; platform CI required
Cambricon	No	N/A	N/A - not affected
Moore	Yes	not run	platform hardware unavailable; platform CI required
Ascend	No	N/A	N/A - not affected

Benchmark / Performance Impact

N/A. This PR relocates the existing kernel for reuse and adds dispatch providers.

Notes for Reviewers

The public schema is unchanged. This PR only shares the existing implementation and fills provider coverage.

InfiniOps API	Alignment target	Evidence
TopkSoftmax(gating_output, bias, is_padding, renormalize, topk_weights, topk_indices, token_expert_indices)	vLLM topk_softmax(topk_weights, topk_ids, token_expert_indices, gating_output, renormalize=False, e_score_correction_bias=None, is_padding=None)	vLLM Python API at pinned commit

InfiniOps groups input tensors, attributes, and explicit outputs according to its C++ operator convention. Thus bias maps to vLLM's e_score_correction_bias, topk_indices maps to topk_ids, and the output tensors are trailing instead of leading. No overload is added.

The NVIDIA result validates the relocated shared kernel and NVIDIA wrapper. Iluvatar, MetaX, and Moore platform CI remain required.