feat(nvidia): add moe_sum operator

[voltjia]

Summary

• Add the NVIDIA moe_sum operator and generated Python binding.
• Support FP32, FP16, and BF16 inputs, optional topk_ids / expert_map routing filters, non-negative strided inputs, non-default streams, and multi-GPU device guards.
• Add focused correctness, descriptor-reuse, routing, stride, stream, and device tests.

Motivation

moe_sum is a representative vLLM MoE primitive needed while moving third-party serving kernels behind InfiniOps. The public name and parameter semantics follow vLLM's stable custom-op schema; InfiniOps keeps its repository-wide convention of input tensors first and the mutable output last.

Alignment reference (pinned to vLLM commit 2f75e7f):

InfiniOps execution API	vLLM schema	Evidence
moe_sum(input, output) moe_sum(input, topk_ids, expert_map, output)	moe_sum(Tensor input, Tensor! output, Tensor? topk_ids=None, Tensor? expert_map=None) -> ()	Stable schema, CUDA semantics, Python wrapper

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

Commit 264501126626b7f3dbf0a05c9cc3ec33dc8a4852 was exported as a clean Git archive and tested on ssh nvidia using accelerator-dev/nvidia:latest (sha256:dd94fce2f83a...), NVIDIA A100-SXM4-80GB, and PyTorch 2.10.0a0+b4e4ee8.

INFINI_RT_ROOT=/opt/infinirt scripts/dev/build.sh nvidia --smoke --jobs 16 --torch-jobs 2 --binding-jobs 2
[build] done

PYTHONPATH=build-nvidia/install python3 -m pytest tests/test_moe_sum.py -q
31 passed, 6 skipped in 21.81s

PYTHONPATH=build-nvidia/install python3 -m pytest tests -m smoke -q
77 passed, 14 skipped, 12331 deselected in 9.63s

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	operator-pruned build and smoke passed	focused moe_sum: 31 passed, 6 skipped; full suite intentionally not run
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

Focused coverage

The focused suite covers both schema overloads, FP32/FP16/BF16 accumulation, int32/int64 routing IDs, negative and remote expert filtering, positive non-contiguous strides, non-default streams, multi-GPU device selection, and safe descriptor reuse. Descriptor reuse permits fresh data pointers with identical metadata and rejects changed shape, stride, dtype, device, or optional-argument presence before launch.

Benchmark / Performance Impact

N/A - this PR establishes the aligned interface and correctness baseline; it makes no performance claim.

Notes for Reviewers

• Review the API mapping and routing semantics against the pinned vLLM schema and CUDA implementation above.
• InfiniOps places mutable output tensors last; this is the only ordering adaptation from the vLLM dispatcher schema.
• Kernel indexing is signed 64-bit. The descriptor rejects negative strides and overflow-prone metadata while preserving non-negative non-contiguous layouts.
• Scope is intentionally limited to one representative NVIDIA MoE operator. moe_align_block_size is being proposed independently.