The Curse of Depth in Large Language Models