What is the significance of multi-head attention in transformer models like GPT and LLaMA?