feat(ascend): add scaled softmax operator

[zhangyue207]

Summary

• Add the ScaledSoftmax base operator for 2D logits/probability normalization.
• Add the Ascend ACLNN implementation using optional aclnnMuls followed by aclnnSoftmax.
• Add focused Ascend tests across shapes, scales, and floating dtypes.

Motivation

ScaledSoftmax is needed as a small inference operator for logits-style normalization on Ascend.

Closes N/A

API Alignment

• Source checked: PyTorch torch.nn.functional.softmax(input, dim=None, _stacklevel=3, dtype=None).
• InfiniOps signature: scaled_softmax(input, scale, out).
• Intentional deviation: scale is explicit because this operator models softmax(input * scale, dim=-1) and uses InfiniOps output-explicit calling style.
• Test coverage: tests/test_scaled_softmax.py compares against torch.nn.functional.softmax(input.to(torch.float32) * scale, dim=-1).

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

Test Results on Supported Platforms

Platform	Built	pytest Result	Notes / Hardware
NVIDIA	N/A	N/A	Not touched.
Iluvatar	N/A	N/A	Not touched.
MetaX	N/A	N/A	Not touched.
Cambricon	N/A	N/A	Not touched.
Moore	N/A	N/A	Not touched.
Ascend	Yes	27 passed	910B4, infiniops-pr-ascend-0512, ASCEND_VISIBLE_DEVICES=0.

Full `pytest` output

python3 -m pip install .[dev] --no-build-isolation -C cmake.define.WITH_ASCEND=ON -C cmake.define.BUILD_ASCEND_CUSTOM=OFF -C cmake.define.AUTO_DETECT_DEVICES=OFF -C cmake.define.GENERATE_PYTHON_BINDINGS=ON
pytest tests/test_scaled_softmax.py --devices ascend -v --tb=short

tests/test_scaled_softmax.py::test_scaled_softmax[...] PASSED

============================== 27 passed in 1.47s ==============================

Benchmark / Performance Impact

N/A

Notes for Reviewers

• This PR intentionally excludes a CPU fallback to keep the Ascend operator PR minimal.
• The Ascend path supports contiguous float16, bfloat16, and float32 tensors.

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits.
• Branch name follows <type>/xxx-yyyy-zzzz.
• Each commit message follows Conventional Commits.
• Small PR is a single squashable commit.
• No stray merge commits from master.
• No fixup! / squash! / wip commits remain.

Scope and Design

• Changes are minimal.
• No dead code, commented-out blocks, debug prints, or ownerless TODO entries.
• No unrelated formatting churn.
• Public API changes are intentional and tested.

General Code Hygiene

• Comments are limited to useful interface notes.
• Added files end with a trailing newline.
• git diff --check passes.
• Identifiers in comments and error messages are wrapped in backticks.
• Comments and error messages are in English.
• Comments and error messages follow project conventions.

C++ Specific

• Code follows the Google C++ style used in this repository.
• clang-format --dry-run --Werror src/base/scaled_softmax.h src/native/ascend/ops/scaled_softmax/kernel.h passes.
• Operator parameter order is inputs, attributes, outputs.
• No exceptions are thrown.
• Error message wording was reviewed.
• Kernel launcher is in kernel.h, matching the local Ascend operator pattern.
• Initializer list order matches member declaration order.
• Blank line rules were reviewed.
• New operator follows the src/base/<op>.h plus src/native/ascend/ops/<op>/ pattern.
• No raw new / delete.

Python Specific

• ruff format --check tests/test_scaled_softmax.py passes.
• ruff check tests/test_scaled_softmax.py passes.
• Test comments/docstrings follow project conventions.
• Type hints are consistent with surrounding tests.

Testing

• Ascend focused test was run and recorded above.
• New functionality has matching tests under tests/.
• Tests use pytest.mark.parametrize and pytest.mark.auto_act_and_assert.
• Default device parameterization is used.
• N/A: No bug fix regression test is needed for this feature PR.

Build, CI, and Tooling

• Project builds through pip install .[dev] in the Ascend test container.
• Python bindings regenerate during the build.
• No new backend or device auto-detection was added.
• No new runtime dependency was added.

Documentation

• N/A: README.md / CONTRIBUTING.md changes are not needed for this small operator addition.
• New public operator behavior is documented in the base header and tests.
• N/A: No user-visible breaking change.

Security and Safety

• No secrets, access tokens, internal URLs, customer data, or personal hardware identifiers have been committed.
• No third-party code was added.
• Pointer and descriptor usage follows existing Ascend AclTensorCache patterns.