DeepSeek V3.2 support and optimization plan

[wuxun-zhang]

To support DeepSeek V3.2 model on XPU, we will add Sparse MLA backend into vllm and several new XPU kernels.

Functionlity

• XPU MLA Sparse backend. Add XPU MLA Sparse backend for DeepSeek v3.2 vllm#33230
• fallback to Triton fp8 moe backend for w8a8 [XPU] Support block fp8 moe by fallback to TritonExpert on XPU vllm#36458
• enable sycl kernels for indexer in vLLM [XPU] Enable topk_per_row and indexer_quant_cache kernels for DeepSeekV3.2 and GLM5 vllm#37888
  • this will depend on next vllm-xpu-kernel release

Optimizations

• add sycl kernel for indexer_k_quant_and_cache. Support indexer_k_quant_and_cache #193
• add sycl kernel for cp_gather_indexer_k_quant_cache. Support cp_gather_indexer_k_quant_cache #210
• add sycl kernel for top_k_per_row_prefill/decode. Add Sycl topk per row kernel #191

[wuxun-zhang]

cc @jikunshang @xinyu-intel @xuechendi