feat(cambricon): link Apex silu_and_mul

[voltjia]

Summary

• Add Cambricon Apex silu_and_mul as implementation slot 16.
• Reuse TorchSiluAndMul<Backend> for tensor wrapping, contiguous staging, and output copy-back.
• Use C10<Device::Type::kCambricon> to bind the caller's CNRT queue to a C10 stream guard.
• Keep the Apex-specific zero-bias and returned-Tensor mapping in ApexSiluAndMul::Call.

Motivation

The Cambricon Apex package exports packed SiLU-and-multiply semantics through:

bias_swiglu_fwd(at::Tensor const&, at::Tensor const&)

Linking that exported ABI avoids copying the vendor kernel into InfiniOps while preserving the existing silu_and_mul(input, out) public interface.

Depends on #875. N/A - no linked issue.

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

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• Hygon (WITH_HYGON)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

Vendor environment:

• host: ssh cambricon
• image: cambricon/pytorch:v1.25.3-torch2.1-anolisos8.8-py310
• image ID: sha256:d3544201bc7039f328abf137d154b24a89a17fdd48a487dc0ae0bdd6d14a5cf6
• linked package: apex==0.1+mlu0.8.0

Runtime evidence collected at 2e781a5b:

INFINI_OPS_OPS=silu_and_mul, WITH_CAMBRICON=ON, WITH_LINKED=ON, WITH_TORCH=OFF
python3 -m pytest tests/test_silu_and_mul.py -m smoke --devices cambricon -q
2 passed, 23 deselected in 0.90s
python3 -m pytest tests/test_silu_and_mul.py --devices cambricon -q
25 passed

At b9ca02c6, both Cambricon main CI and Cambricon shadow CI passed. Current head d0eac65d only renames the logical mapping from apex to apex_glu_activation and updates its binding. Build, formatting, Ruff, and the main Cambricon/Ascend/Moore/MetaX jobs pass; the remaining self-hosted jobs are pending.

Test Results on Supported Platforms

Focused host-independent validation at c2bbc2df:

python -m pytest --noconftest tests/test_resolve_linked_ops.py tests/test_generate_wrappers.py -q
39 passed
ruff format --check tests/test_silu_and_mul.py
1 file already formatted
ruff check tests/test_silu_and_mul.py
All checks passed!
clang-format --dry-run --Werror <Cambricon linked C++ files>
exit 0

Benchmark / Performance Impact

N/A - this PR validates correctness and linkage; it makes no performance claim.

Notes for Reviewers

• ApexSiluAndMul inherits C10<Device::Type::kCambricon> and supplies only the external ABI call.
• The logical library name apex_glu_activation identifies the concrete glu_activation*.so DSO; other Apex DSOs use separate mappings, such as apex_fused_layer_norm.
• C10<Device::Type::kCambricon> uses torch_mlu::getStreamFromExternal and c10::StreamGuard, so execution stays on the caller-provided stream.
• Apex represents no bias with a zero-length integer Tensor and returns a Tensor; ApexSiluAndMul::Call maps those details to the shared output-oriented adapter.
• WITH_TORCH=OFF disables the generated ATen backend. The linked implementation still uses the installed Torch/torch-mlu C++ ABI required by the external library.
• The Cambricon environment requires fix(linked): support readelf without demangling #875 because its GNU readelf does not provide --demangle.