Understanding and Coding the Self-Attention Mechanism of LLMs from Scratch