[Triton] [ATOM] MXFP8 GEMM and A8W4 MOE optimization

[k50112113]

This PR is co-authored by @azaidy and @k50112113

This PR includes:

1. Added MXFP8 GEMM (non-shuffled and pre-shuffled), tunned with DSV4 shapes
2. Added all other mxfp8 GEMM related kernels, including: per_1x32_mxfp8_quant_triton, rmsnorm_mxfp8_quant, dual_rmsnorm_mxfp8_quant, fused_flatten_mxfp8_quant.
3. A8W4 MOE fused with activation and downcast between the 2 MOEs
4. Optimized Topk and routing for the A8W4 MOE
5. Fix fused_flatten_fp8_group_qaunt bug when N is not power of 2

[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 3292 --add-label <label>