FlashAttention 3

Fast, memory-efficient exact attention — 2× speedup for training and inference on H100.