feat(cambricon): add topk softmax provider

[baominghelly]

Summary

• Add a native Cambricon slot 0 provider for topk_softmax in src/native/cambricon/ops/topk_softmax/.
• Support float16, bfloat16, and float32 gating inputs with int32, uint32, or int64 expert indices.
• Extend the existing native tests to run on Cambricon while preserving the explicit NVIDIA linked slot 16 tests.

Motivation

Cambricon currently has no native topk_softmax provider for MoE routing. This change adds the slot 0 implementation so gating probabilities, expert selection, optional bias, padding sentinels, and optional renormalization can execute on MLU.

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

Targeted Cambricon operator build and test were run. The repository smoke suite was not run, so this PR remains Draft.

cmake -S . -B /workspace/build/retest-v3-topk_softmax \
  -DAUTO_DETECT_DEVICES=OFF \
  -DWITH_CPU=ON \
  -DWITH_CAMBRICON=ON \
  -DWITH_TORCH=OFF \
  -DWITH_LINKED=OFF \
  -DGENERATE_PYTHON_BINDINGS=ON \
  -DINFINI_OPS_OPS=topk_softmax \
  -DINFINI_RT_ROOT=/workspace/install/infinirt-master-test \
  -DPython_EXECUTABLE=/torch/venv3/pytorch/bin/python3.10 \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_PREFIX=/workspace/build/retest-v3-topk_softmax-site/infini
cmake --build /workspace/build/retest-v3-topk_softmax --parallel "$(nproc)"
cmake --install /workspace/build/retest-v3-topk_softmax
PYTHONNOUSERSITE=1 \
PYTHONPATH=/workspace/build/retest-v3-topk_softmax-site \
LD_LIBRARY_PATH=/workspace/build/retest-v3-topk_softmax-site/infini:/workspace/install/infinirt-master-test/lib:/usr/local/neuware/lib64:/usr/local/neuware/lib \
/torch/venv3/pytorch/bin/python3.10 -m pytest -q \
  tests/test_topk_softmax.py --devices cambricon --maxfail=1

[100%] Built target ops
32 passed, 51 skipped in 0.45s

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	No	N/A - not affected	Existing linked slot 16 tests remain unchanged
Iluvatar	No	N/A - not affected	N/A - not affected
MetaX	No	N/A - not affected	N/A - not affected
Cambricon	Yes	Targeted build passed	tests/test_topk_softmax.py: 32 passed, 51 skipped
Moore	No	N/A - not affected	N/A - not affected
Ascend	No	N/A - not affected	N/A - not affected

Full `pytest` output (optional)

........ssss........ssss........ssss......s..sssssssssssssssssssssssssss [ 86%]
sssssssssss                                                              [100%]
32 passed, 51 skipped in 0.45s

Benchmark / Performance Impact

N/A - no performance benchmark was run.

Notes for Reviewers

• Native tests continue to use the default slot 0. NVIDIA linked coverage continues to explicitly select slot 16.
• Selection uses optional bias, but returned weights are gathered from the original softmax probabilities. Ties select the smaller expert index.
• Padding writes the all-ones sentinel for index outputs. Cambricon uint32 cases are skipped because torch_mlu cannot construct the required uint32 test tensors.
• Per-task scratch storage must fit in NRAM; the launcher checks the required size before launch.
• Validation is limited to the targeted Cambricon build and operator test file; the repository smoke suite and full suite were not run.