fix(fused_moe): restore gfx942 FLAT kernel support

[bingxche]

Summary

• restore gfx942 to the FLAT FMoE architecture allowlist while retaining the unsupported-architecture guard
• allow GLM-5.x FP8 low-token decode to load the shipped gfx942 flat_pf3 kernel again

Root cause

AITER #4040 added the gfx942 FLAT kernel, tuned GLM rows, and the gfx942/gfx950 allowlist. #4394 later narrowed only the guard back to gfx950, while leaving the gfx942 kernel and config in place. SGLang GLM-5.1 therefore selects the valid gfx942 kernel but fails before launch.

Observed in both SGLang jobs:

• https://github.com/sgl-project/sglang/actions/runs/30716848145/job/91730352050
• https://github.com/sgl-project/sglang/actions/runs/30716848145/job/91730351399

Test plan

• git diff --check
• Run the targeted gfx942 FLAT FMoE case for M=4/8, model_dim=6144, inter_dim=256, E=257, topk=9
• Rerun SGLang nightly-8-gpu-glm51
• Rerun SGLang nightly-8-gpu-glm51-rocm720

Local Python/ROCm tests were not run because the authoring environment is Windows without Python or AMD GPUs.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 4628 --add-label <label>

[bingxche]

Validation runs for AITER head dc127de9d1fb63349477dd1448951c32c50db563:

• SGLang nightly-8-gpu-glm51 (ROCm 7.0): https://github.com/sgl-project/sglang/actions/runs/31170483868/job/92840966809
• SGLang nightly-8-gpu-glm51-rocm720 (ROCm 7.2): https://github.com/sgl-project/sglang/actions/runs/31170484511/job/92840969922

Both workflows were dispatched from SGLang main with only the targeted GLM-5.1 job selected and continue_on_error=false.

[bingxche]

Hi @yadaish, could you please review? Thanks!