Self-Boosting Large Language Models with Synthetic Preference Data