feat(cambricon): add gelu provider

[baominghelly]

Summary

• Add a native Cambricon GELU provider in src/native/cambricon/ops/gelu/.
• Use CNNL for exact GELU on float16 and float32, with a custom MLU kernel for tanh approximation and bfloat16.
• Support contiguous and strided input/output tensors, in-place execution, scalar tensors, and empty tensors.
• Skip Cambricon float64 cases in tests/test_gelu.py because CNNL GELU does not support that dtype.

Motivation

InfiniOps did not have a native Cambricon implementation for the canonical gelu operator. This provider enables GELU execution on Cambricon while retaining CNNL where it is reliable and using a custom fallback for the remaining supported paths.

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 after rebasing onto origin/master. The repository-wide smoke suite was not run.

cmake -S . -B /workspace/build/pr-gelu-20260901 \
  -DAUTO_DETECT_DEVICES=OFF \
  -DWITH_CPU=ON \
  -DWITH_CAMBRICON=ON \
  -DWITH_TORCH=OFF \
  -DWITH_LINKED=OFF \
  -DGENERATE_PYTHON_BINDINGS=ON \
  -DINFINI_OPS_OPS=gelu \
  -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/pr-gelu-20260901-site/infini
cmake --build /workspace/build/pr-gelu-20260901 --parallel "$(nproc)"
cmake --install /workspace/build/pr-gelu-20260901
PYTHONNOUSERSITE=1 \
PYTHONPATH=/workspace/build/pr-gelu-20260901-site \
LD_LIBRARY_PATH=/workspace/build/pr-gelu-20260901-site/infini:/workspace/install/infinirt-master-test/lib:${LD_LIBRARY_PATH:-} \
/torch/venv3/pytorch/bin/python3.10 -m pytest -q tests/test_gelu.py --devices cambricon --maxfail=1

42 passed, 14 skipped in 0.35s

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	No	N/A - not affected	N/A - not affected
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_gelu.py: 42 passed, 14 skipped; full suite not run
Moore	No	N/A - not affected	N/A - not affected
Ascend	No	N/A - not affected	N/A - not affected

Full `pytest` output (optional)

ssssssssssssss..........................................                 [100%]
42 passed, 14 skipped in 0.35s

Benchmark / Performance Impact

N/A — no performance benchmark was run.

Notes for Reviewers

• Exact GELU for float16 and float32 uses CNNL.
• Tanh-approximate GELU and bfloat16 use the custom GeluUnion kernel. The custom path computes in float32 before converting back to the output dtype.
• The custom path supports contiguous and non-contiguous tensors and allocates workspace only for copied shape/stride metadata.
• float64 is intentionally unsupported on Cambricon and is skipped in the MLU test cases.
• Please review the numerical behavior of the custom tanh and bfloat16 paths. The full repository smoke/full suite and performance benchmarks were not run.