refactor(ops): migrate softmax to canonical InfiniOps API

[voltjia]

Summary

• migrate the InfiniCore softmax adapter from deprecated SoftmaxInfinilm to canonical Softmax
• preserve the existing input, dimension, optional dtype, and output mapping
• keep the InfiniCore public API, graph planning, and execution behavior unchanged

API alignment

InfiniCore adapter	Previous InfiniOps API	Canonical InfiniOps API	Alignment target and evidence
Softmax	SoftmaxInfinilm::Call(handle, config, input, dim, dtype, out)	Softmax::Call(handle, config, input, dim, dtype, out)	PyTorch Python API: torch.nn.functional.softmax(input, dim=None, _stacklevel=3, dtype=None). InfiniOps omits Python's internal _stacklevel argument and adds the explicit trailing C++ output tensor.

Dependency

This PR is based on refactor/migrate-paged-caching-infiniops / #1467 and advances the InfiniOps gitlink from e733e325 to eec65a13, the head of InfiniOps #890, which adds canonical CUDA-compatible softmax providers.

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

Validation

Validated commit 86bec726 remotely on ssh nvidia with accelerator-dev/nvidia:latest, using the exact InfiniOps #890 head eec65a13 in a temporary integration merge:

• built and installed _infinicore with the canonical softmax provider in the wrapper allowlist
• compiled and ran a temporary C++ NVIDIA smoke test through infinicore::op::softmax; output matched a row-wise host reference
• 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Superseded by #1480, which consolidates the canonical InfiniOps adapter migrations while preserving each logical change as a separate commit.