feat(cuda): add canonical softmax providers

[voltjia]

Summary

• Add canonical Softmax CUDA providers for NVIDIA, Iluvatar, MetaX, and Moore.
• Reuse one generic CUDA implementation and add focused coverage across dimensions, dtypes, and non-contiguous layouts.

Motivation

InfiniCore still uses the deprecated SoftmaxInfinilm compatibility operator. Canonical provider coverage is required before that adapter can migrate to the PyTorch-aligned Softmax API.

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_softmax.py --devices nvidia -q
24 passed

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	build passed	24 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. The canonical providers reuse one generic implementation.

Notes for Reviewers

The public schema already existed; this PR adds native provider coverage and the metadata needed by the generic implementation.

InfiniOps API	Alignment target	Evidence
Softmax(input, dim, dtype, out)	PyTorch torch.nn.functional.softmax(input, dim=None, _stacklevel=3, dtype=None)	PyTorch softmax documentation

InfiniOps omits Python's internal _stacklevel argument and uses an explicit trailing output tensor for its C++ API. This PR does not add overloads or remove SoftmaxInfinilm. The vendor backend headers are thin wrappers around src/native/cuda/ops/softmax/kernel.h.

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