What is FlashAttention, and how does it reduce memory reads and writes during the self-attention step?
