Usual implementation of attention transformers (SDPA) is kind of bad, actually

**Brief Summary** The project highlights that the standard scaled‑dot‑product attention (SDPA) implementation used in most transformer libraries is sub‑optimal,