feat(intel-xpu): support MiniMax-H3 Turbo-SLA inference

[qiuxin2012]

Summary

Add Intel XPU support for MiniMax-H3 Turbo-SLA inference.

Changes

• Add XPU implementations for SLA block-map generation and sparse-block attention.
• Expose the following sycl_kernels APIs:
  • sla_block_map
  • sparse_block_attention
  • sla_sparse_attention
• Integrate intel_xpu_cute_attn into dynamic_sparse_attn.
• Add MiniMax-H3 Turbo-SLA Intel XPU configuration and run script.
• Support dynamic application of the Turbo-SLA LoRA weights.
• Use dense CUTE attention for the refiner and SLA sparse attention for the main transformer.