Multi-Head Attention

Explain how attention heads divide the model dimension, how the projection and reshape operations work, and how variants change KV-cache cost.