feat(linked): add MetaX vLLM silu_and_mul backend

[voltjia]

Summary

• Add src/linked as a peer implementation backend for operators recovered from installed shared libraries.
• Support multiple provider implementations for one (operator, device) through provider-named YAML/C++ files and independent implementation slots.
• Organize linked implementations under src/linked/<transport>/<device> and add MetaX vLLM silu_and_mul at implementation slot 16.

Motivation

Some accelerator vendors distribute optimized vLLM operators only as shared objects. InfiniOps needs a narrow, auditable integration path for those implementations without vendoring generated kernels, adding a submodule, or treating them as native source implementations.

The public API remains InfiniOps silu_and_mul(input, out), with packed-input semantics following vLLM SiluAndMul. The installed MetaX provider ABI is verified independently by nm and readelf as:

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

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

Environment: ssh metax, image cr.metax-tech.com/public-ai-release/maca/vllm-metax:0.10.2-maca.ai3.2.1.7-torch2.6-py310-ubuntu22.04-amd64 (sha256:6e519687a9e4...), exact commit 20ceb1cc50f575449ba258a746f736e3bd6327f9.

INFINI_OPS_OPS=silu_and_mul, WITH_METAX=ON, WITH_LINKED=ON, WITH_TORCH=OFF
python -m pytest tests -m smoke --devices metax -q
46 passed, 40 skipped, 4269 deselected

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Focused Result / Notes
MetaX	Yes	Operator-pruned linked build and smoke passed	tests/test_silu_and_mul.py: 50 passed
Other platforms	No	N/A	Added by separate stacked PRs

Additional validation

tests/test_resolve_linked_ops.py + tests/test_generate_wrappers.py
38 passed
libinfiniops.so NEEDED: vllm/_C.abi3.so, torch, c10, and MetaX runtime libraries
RUNPATH: $ORIGIN:<torch/lib>:<vllm>:<MetaX runtime lib>

Benchmark / Performance Impact

N/A - this PR establishes an integration path and makes no performance claim.

Notes for Reviewers

• Linked providers live under src/linked/<transport>/<device>. Platform-level <library>.yaml files discover DSOs; operator-level provider files use one stem, for example vllm.yaml, vllm.h, and vllm.cc.
• Platform library YAML contains python_distribution_package and library_glob for DSO discovery.
• Operator binding YAML identifies its DSO with library: vllm; multiple implementations may reference the same DSO.
• Linked implementations use unique slots from 16; kMaxImplementations is 32.
• TorchSiluAndMul<Backend> owns tensor wrapping, contiguous staging, and output copy-back. Each provider Backend inherits C10<device> for device identity and stream handling, then adds only its exact Call ABI.
• The shared Torch adapter declaration and implementation live together in src/linked/torch/ops/silu_and_mul.h. Provider .cc remains the ABI and explicit-instantiation boundary; provider .h uses extern template so generated translation units do not instantiate it again.
• C10<device> converts the caller's external stream to the vendor Torch stream type and exposes the matching stream guard; it does not replace the caller-selected stream.
• The provider DSO is linked but not bundled. The resolver rejects ambiguous libraries, missing symbols, and the same global C++ symbol exported by multiple loaded DSOs.
• Loading multiple provider DSOs that export the same global C++ symbol is unsupported because of ELF symbol interposition.