FlashPrefill V2: 47x Speedup for Long-Context LLM Prefill

FlashPrefill V2 introduces block-sparse prefill attention, achieving up to 47x speedup over FlashAttention-2 on H20 GPUs. Includes mean correction term, FP8 support, paged KV cache, and SGLang integration — a production-ready breakthrough for efficient long-context serving.