Unleashing LLMs in Bayesian Optimization: Preference-Guided Framework for Scientific Discovery