[Feat] Fused qknorm + quant for dpsk v2 model

[qichu-yun]

Motivation

Improve DeepSeek V2 performance by enabling fused q/k RMSNorm + quantization for both FP8 and MXFP4 paths.

Technical Details

This commit enables fuse_qknorm_quant for MXFP4, preserves unshuffled MXFP4 attention projection weights for raw-scale GEMM, and routes SGLang q_b_proj through the proper FP4 GEMM path when fused quantization is used. Existing non-fused and fallback paths remain unchanged.

Test Result

before：

after:

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[valarLip]

we should better remove e8m0_shuffle by other fusion

[qichu-yun]

This e8m0_shuffle() is only on the fallback path. The native SGLang-style FP4 GEMM path consumes raw q_scale directly and returns before this line, so the optimized path has no post-shuffle. It is only used when we fall back to ATOM q_b_proj, which still requires shuffled scale layout. Removing it without removing/replacing the fallback would reintroduce the scale-layout correctness issue.