feat(moore): link vLLM silu_and_mul

[voltjia]

Summary

• Add the Moore vLLM mapping for silu_and_mul(at::Tensor&, at::Tensor&) as implementation slot 16.
• Name the implementation vllm.yaml, vllm.h, and vllm.cc, reusing the shared linked Torch silu_and_mul template.
• Bridge the caller-provided MUSA stream through C10<Device::Type::kMoore>.

Motivation

Moore vLLM exposes an optimized silu_and_mul implementation through a C++ ABI. InfiniOps links it to the public silu_and_mul(input, out) API. C10<Device::Type::kMoore> owns device and stream integration, while VllmSiluAndMul owns the exact external ABI call.

Semantics follow vLLM SiluAndMul. The pinned Moore image exports:

silu_and_mul(at::Tensor&, at::Tensor&)

Depends on #873. 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

Manual runtime evidence at 44413c861c55d825da4b46d3a4d716f8391b96b2: ssh huangjiacheng@172.22.162.98, image zx_vllm_musa_436:vllm017-cli-ok-mate-disabled (sha256:0431d022137d...), Torch 2.7.1, vLLM 0.17.1.dev0+gb31e9326a.d20260615.empty.

INFINI_OPS_OPS=silu_and_mul, WITH_MOORE=ON, WITH_LINKED=ON, WITH_TORCH=OFF
python3 -m pytest tests/test_silu_and_mul.py -m smoke --devices moore -q
4 passed, 46 deselected

At current head d70f191b, both Moore main CI and Moore shadow CI pass.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Focused Result / Notes
Moore	Yes	Manual operator-pruned smoke passed; current-head main and shadow CI pass	tests/test_silu_and_mul.py: 50 passed at 44413c86
MetaX	No	N/A	Provided by dependency #873
Other platforms	No	N/A	Not affected

Additional validation

tests/test_resolve_linked_ops.py + tests/test_generate_wrappers.py
39 passed
NEEDED: vLLM _C, libtorch_cpu.so, libc10.so, libmusart.so.4, libmublas.so.1
RUNPATH: $ORIGIN:<torch/lib>:<vllm>

Benchmark / Performance Impact

N/A - this PR changes integration behavior and makes no performance claim.

Notes for Reviewers

• Moore implementation files use the vllm.* stem. Platform library YAML contains the Python package and DSO glob.
• TorchSiluAndMul<Backend> provides tensor conversion, staging, and copy-back. VllmSiluAndMul::Call contains the exact external ABI call.
• C10<Device::Type::kMoore> uses c10::musa::getStreamFromExternal and c10::StreamGuard. The guard applies the caller-selected device and stream.
• The vendor compatibility header <c10/musa/MUSAMacros.h> remains required because that image's MUSAStream.h does not define C10_MUSA_API itself.
• The non-default-stream test is included in the 50 passed result.