Bayesian Optimization with LLM-Based Acquisition Functions for Natural Language Preference Elicitation