feat(nvidia): link vLLM fused_marlin_moe

[voltjia]

Summary

• Add the NVIDIA linked slot-16 implementation of vLLM fused_marlin_moe with one complete upstream-aligned overload and InfiniOps' trailing out convention.
• Compose the provider's registered _moe_C::moe_align_block_size and _moe_C::moe_wna16_marlin_gemm operators with ATen SiLU/multiply and reduction, without copying provider kernels.
• Cover out-of-place, in-place aliasing, expert mapping, empty activation-order metadata, non-default streams, and linked capability gates in the existing single operator test module.

Motivation

On Ampere, vLLM uses fused_marlin_moe as a complete quantized MoE path. Its public operator is registered by Python as a composite rather than exported by the provider DSO, so InfiniOps cannot honestly bind a same-named shared-library symbol. This PR reconstructs the pinned composite from the two exact Dispatcher contracts supplied by the installed vLLM _moe_C library and standard ATen operations.

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)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

Remote environment: A100-SXM4-80GB, vllm/vllm-openai:v0.10.0 (vLLM 0.10.1.dev1+gbcc0a3cbe, PyTorch 2.7.1+cu128), CUDA 12.8, CUDA-12-compatible InfiniRT.

# Clean Release/NDEBUG, SM80, WITH_LINKED=ON, smoke union plus fused_marlin_moe
linked resolver tests: 32 passed
tests/test_fused_marlin_moe.py: 7 passed
pytest tests -m smoke -q: 78 passed, 22 skipped, 2 failed, 7809 deselected

The two smoke failures are the same float32 GEMM tolerance cases already reproduced on a fresh origin/master build in the identical environment while validating the preceding Marlin primitive PR. No additional smoke failure remains after running from the correct generated-header working directory.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Clean Release linked build and focused operator tests passed; smoke matched the established master baseline	Full suite not run; validation was intentionally operator-pruned
Iluvatar	No	N/A - not affected	N/A
MetaX	No	N/A - not affected	N/A
Cambricon	No	N/A - not affected	N/A
Moore	No	N/A - not affected	N/A
Ascend	No	N/A - not affected	N/A

Additional focused validation

clang-format 21.1.8 dry run: passed
focused strict wrapper generation: passed
real provider schemas and CUDA dispatch kernels: passed
FP4, global_scale, and undersized workspace rejection: passed
out-of-place, in-place alias, identity expert_map, empty [0] act-order, and non-default stream: passed
git diff --check: passed

Benchmark / Performance Impact

N/A - this PR establishes the composite interface and correctness path and makes no performance claim.

Notes for Reviewers

• The public argument order follows vLLM commit bcc0a3cb, with the InfiniOps-required output moved to the end.
• vLLM's high-level vllm::fused_marlin_moe appears only after Python registration; it is not a DSO contract. The binding therefore declares the exact low-level align and Marlin schemas that its C++ source actually calls.
• inplace=true requires out to alias hidden_states; inplace=false requires a distinct output. The linked boundary checks this in Release builds.
• The pinned high-level qtype set is enforced independently from the broader low-level primitive. FP4 and global scales are rejected before tensor conversion because InfiniRT cannot currently represent the required float8 scales.
• This PR is stacked on fix(linked): support multiple operator schemas #913, which extends the existing operator_schema YAML field to validate both low-level Dispatcher contracts from one DSO.