refactor(ops): use canonical InfiniOps softmax API

[voltjia]

What

• Replace SoftmaxInfinilm with the canonical InfiniOps Softmax API.
• Preserve the existing InfiniCore axis and output handling.

Alignment

InfiniCore wrapper	InfiniOps call	Alignment basis
softmax(input, dim, out)	Softmax(input, dim, dtype=nullopt, out)	torch.nn.functional.softmax(input, dim, dtype=None), InfiniOps Softmax, InfiniOps #890

The omitted dtype remains std::nullopt, matching the existing behavior of preserving the input dtype.

Why

InfiniOps now exposes the standard PyTorch-aligned softmax interface and its CUDA provider, so the deprecated InfiniLM compatibility class is no longer needed by this adapter.

Scope

This is stacked on #1476 for the validated InfiniOps pin. It does not change the public InfiniCore softmax API or add overloads.

Screenshots: N/A (backend adapter migration only).

Validation

Run on ssh nvidia in accelerator-dev/nvidia:latest on NVIDIA A100 GPUs:

• Full infinicore_cpp_api and _infinicore build/install passed.
• Python extension import and dynamic linking passed.
• python3 scripts/format.py --ref 850fb3f7 --path src --check with clang-format 16.0.6.
• git diff --check.