The I/O Complexity of Attention, or How Optimal is Flash Attention?