Qwen3.5 support and optimization plan

[wuxun-zhang]

We would like to support Qwen3.5 model from functionality perspective firstly and then further optimize kernel performance to improve E2E performance.

Functionality

• add XPU gdn op support to vLLM [XPU] Support Qwen3-next/Qwen3.5 vllm#33657
  • this should be updated and merged after refactor PR [Model] Extract GatedDeltaNetAttention into shared layer for Qwen3Next and Qwen3.5 vllm#37975
• support fp32 ssm_state in chunk_fwd_o kernel Support f32 ssm_state in GDN kernel for Qwen3.5 #220

Performance optimizations

GDN attention

Base kernel version: #156

• optimize l2norm kernel Optimize l2norm in GDN kernel for Qwen3.5 #222
• optimize chunk_fwd_o kernel
• optimize grouped gemm kernel

Layer Norm

• add sycl kernel for GemmaRMSNorm and RMSNormGated support gemma_rms_norm and rms_norm_gated kernels #214

Qwen3 VisionTransformer

(placeholder)

[wuxun-zhang]

cc @xinyu-intel

[xinyu-intel]

consider #177, I think the group gemm kernel need to be optimized for Qwen workloads.