Flash attention is not faster than vanilla attention

M2 Pro 10c 32G

<img width="2464" height="1176" alt="Image" src="https://github.com/user-attachments/assets/8c5c20b8-a89b-44ab-a896-7c6d20c83c67" />

Bottleneck is somewhere else in the current implementation. Needs some profiling.