Add documentation for standard attention layer principles

This document explains the principles and calculations of standard attention layers, including the computation flow, projection of inputs to Q/K/V, rotary position encoding (RoPE), multi-head attention (MHA), and Flash Attention techniques.