fix(kernels): catch C++ exceptions at the FFI boundary and surface the message

[FeathBow]

Description

Fixes #520

FlashInfer can throw std::exception from host-side dispatch before any CUDA kernel launch. When that exception crosses an extern "C" boundary, the process aborts without surfacing the message. This was hit by Qwen3-14B: 40/8 heads -> GQA group size 5, which is not instantiated by DISPATCH_GQA_GROUP_SIZE.

• Add csrc/shared/ffi_guard.cuh + ffi_last_error.cu: guarded FFI entries catch C++ exceptions at the boundary, store what() in a thread_local, clear stale messages on entry, and return a -1 sentinel.
• Rust wrappers append the stored message on -1; ffi_exception_message is public for direct-FFI callers outside the kernels crate. The qwen35 hd256 prefill path and qwen3 kernel bench now append the message too.
• SUPPORTED_GQA_GROUP_SIZES (&[1, 2, 3, 4, 8], matching the FlashInfer dispatch table) is exported from the kernels crate; qwen3 config load rejects unsupported group sizes up front.
• glm52 top-k now returns i32 at the FFI declaration so the -1 sentinel cannot materialize an invalid CUresult.

The new FFI test drives decode at group 7 and asserts -1 plus a message naming the group size, instead of a process abort.

Test Env

Single GPU (sm_89, x86_64):

• FFI test green.
• Qwen3-4B (group 4) loads and serves unchanged.
• Qwen3-14B now fails at load with unsupported GQA group size 40/8; the attention kernels instantiate [1, 2, 3, 4, 8].
• kimi-k2 and glm52 feature builds compile clean with the guard additions (cargo check targeting sm_90 with the full moe dependency chain).

Type of Change

• Bug fix (non-breaking change which fixes an issue)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 154764075e

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".