Enhancing High-Quality Code Generation in Large Language Models with Comparative Prefix-Tuning