feat(kernels): gate MoE third-party substrates

[xiaguan]

Summary

• add DeepGEMM and FlashMLA as kernel submodules and introduce the shared moe feature for DeepEP/DeepGEMM/FlashMLA substrates
• add the narrow glm52 kernel surface for DeepGEMM scale layout/grouped FP8 contracts and FlashMLA SM90 sparse decode wrappers
• update kernel docs and keep non-GLM52/router/indexer/PP/TRTLLM surfaces out until the model crate has stable callers
• align bench_serving with the current Qwen3 engine signature so the existing pre-commit clippy gate passes

Verification

• cargo fmt
• OPENINFER_CUDA_SM=90 cargo check -p openinfer-kernels --no-default-features
• OPENINFER_CUDA_SM=90 OPENINFER_NCCL_ROOT=/data/code/workspace-rustllm/ep-moe-demo/.venv/lib/python3.12/site-packages/nvidia/nccl cargo check -p openinfer-kernels --no-default-features --features moe
• OPENINFER_CUDA_SM=90 OPENINFER_NCCL_ROOT=/data/code/workspace-rustllm/ep-moe-demo/.venv/lib/python3.12/site-packages/nvidia/nccl cargo check -p openinfer-kernels --no-default-features --features glm52
• OPENINFER_CUDA_SM=90 OPENINFER_NCCL_ROOT=/data/code/workspace-rustllm/ep-moe-demo/.venv/lib/python3.12/site-packages/nvidia/nccl cargo check -p openinfer-kernels --no-default-features --features kimi-k2
• OPENINFER_CUDA_SM=90 cargo check --release -p openinfer-server --bin bench_serving
• pre-commit hooks via OPENINFER_CUDA_SM=90 OPENINFER_NCCL_ROOT=/data/code/workspace-rustllm/ep-moe-demo/.venv/lib/python3.12/site-packages/nvidia/nccl git commit ...

Notes

• GLM5.2 FlashMLA sparse decode is SM90-only, fixes V32 topk=2048, and intentionally does not expose dynamic topk_length.
• The grouped DeepGEMM compute entry is fail-closed with CUDA_ERROR_NOT_SUPPORTED until a real DeepGEMM runner is wired.
• Upstream FlashMLA CHECK_CUDA / FLASH_ASSERT paths can still terminate the process on internal CUDA/launch failures; this PR only converts C++ assertion exceptions at the FFI boundary into CUresult.
• No performance win is claimed; this is feature-gated substrate/API bring-up, so no A/B benchmark is attached.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7ee9da3705

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Avoid making Kimi require unused GLM submodules

When building with --features kimi-k2, this feature edge now enables moe, and build.rs::require_moe_submodules asserts that both DeepGEMM and FlashMLA are checked out before any Kimi CUDA is compiled. The Kimi/DeepEP paths do not include those repositories, so existing Kimi-only checkouts that initialized only DeepEP will fail even though the model does not use the GLM5.2 wrappers; please split the DeepEP substrate from the GLM5.2 DeepGEMM/FlashMLA checks or require those only under glm52.

Useful? React with 👍 / 👎.