refactor(ops): migrate topksoftmax to canonical InfiniOps API

[voltjia]

Summary

• migrate the InfiniCore topksoftmax adapter from deprecated TopkSoftmaxInfinilm to canonical TopkSoftmax
• provide the canonical auxiliary token_expert_indices output as plan-owned internal storage
• keep the InfiniCore public two-output API and existing norm behavior unchanged

API alignment

InfiniCore adapter	Previous InfiniOps API	Canonical InfiniOps API	Alignment target and evidence
Top-k softmax	TopkSoftmaxInfinilm::Call(handle, config, gating_output, topk, renormalize, topk_weights, topk_indices)	TopkSoftmax::Call(handle, config, gating_output, bias, is_padding, renormalize, topk_weights, topk_indices, token_expert_indices)	vLLM Python API: topk_softmax(topk_weights, topk_ids, token_expert_indices, gating_output, renormalize=False, e_score_correction_bias=None, is_padding=None). InfiniOps groups inputs, attributes, and explicit outputs; bias maps to e_score_correction_bias, and topk_indices maps to topk_ids.

The existing InfiniCore public API does not expose bias, padding, or the auxiliary token-to-expert output. This adapter passes null optional inputs and allocates the required third output in plan-owned storage. Its shape retains the caller's existing topk selection, while the canonical operator derives topk from the output shape.

Dependency

This PR is based on refactor/migrate-paged-caching-infiniops / #1467 and advances the InfiniOps gitlink from e733e325 to 1a712d0c, the head of InfiniOps #891, which extends the canonical provider to the CUDA-compatible backends used by InfiniCore.

Against #1467, this PR changes one adapter file and the InfiniOps gitlink.

Validation

Validated commit 3199b309 remotely on ssh nvidia with accelerator-dev/nvidia:latest, using the exact InfiniOps #891 head 1a712d0c in a temporary integration merge:

• built and installed _infinicore with the canonical topk_softmax provider in the wrapper allowlist
• ran python3 test/infinicore/ops/topksoftmax.py --nvidia: 24/24 passed
• clang-format 16.0.6 --dry-run --Werror passed for the changed C++ file
• git diff --check passed

[voltjia]

Superseded by #1480, which consolidates the canonical InfiniOps adapter migrations while preserving each logical change as a separate commit.