fix(cambricon): prevent duplicate half definitions

[baominghelly]

Summary

• Pre-include a Cambricon host compatibility header that gives Neuware FP16/BF16 helper definitions standard C++ inline linkage when they are parsed by GNU/Clang host compilers.
• Apply the compatibility include to both the infiniops library and generated Python ops extension, with a regression test covering the CMake configuration.

Motivation

Neuware 6.0 leaves the BANG host annotation empty in its FP16/BF16 headers when they are parsed by a host C++ compiler. When generated dispatch or binding sources are split across multiple translation units, each object emits strong definitions for the same __half/BF16 helpers and the final shared-library link fails with multiple definition errors.

This fixes the linkage at the vendor-header boundary without changing dispatch batching or forcing all generated operators into one translation unit.

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

PYTHONNOUSERSITE=1 python -m pip install . --target /workspace/build/pr-c412dbb-site \
  --no-build-isolation --no-deps --no-cache-dir \
  -C cmake.define.INFINI_RT_ROOT=/workspace/install/infinirt-master-test \
  -C cmake.define.AUTO_DETECT_DEVICES=OFF \
  -C cmake.define.AUTO_DETECT_BACKENDS=OFF \
  -C cmake.define.WITH_CPU=ON \
  -C cmake.define.WITH_CAMBRICON=ON \
  -C cmake.define.WITH_NVIDIA=OFF \
  -C cmake.define.INFINI_OPS_SMOKE_BUILD=ON

Successfully built InfiniOps
Successfully installed InfiniOps-0.1.0

PYTHONNOUSERSITE=1 PYTHONPATH=/workspace/build/pr-c412dbb-site \
  python -m pytest tests -m smoke -q --devices cambricon

1 failed, 77 passed, 38 skipped, 4635 deselected in 5.91s

The single failure is the existing Cambricon CNNL FP16 GEMM numerical-tolerance case in tests/test_gemm.py (rtol=0.001, atol=0.001). It is outside the files and linkage path changed by this PR.

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	Wheel build passed; smoke: 77 passed, 38 skipped, 1 existing GEMM tolerance failure	Full suite not run; wrapper regression tests: 40 passed
Moore	No	N/A - not affected	N/A - not affected
Ascend	No	N/A - not affected	N/A - not affected

Additional validation

clang-format version 21.1.2
clang-format --dry-run --Werror src/native/cambricon/host_bang_compat.h
# passed

git diff --check origin/master...HEAD
# passed

python -m pytest -q tests/test_generate_wrappers.py
40 passed in 5.76s

Benchmark / Performance Impact

N/A. This changes host-side linkage semantics during compilation and does not change operator kernels or runtime dispatch behavior.

Notes for Reviewers

• The forced include is gated by WITH_CAMBRICON, GNU/Clang, and host C++ compilation. It does not apply to .mlu device compilation or other backends.
• Both infiniops and the generated Python ops target need the compatibility include because each can contain multiple generated dispatch translation units.
• This intentionally keeps the normal dispatch batch size and does not add INFINI_OPS_DISPATCH_BATCH_SIZE.
• The known Cambricon FP16 GEMM tolerance failure is intentionally not mixed into this focused linkage fix.