Preference Tree Optimization: Enhancing Goal-Oriented Dialogue with Look-Ahead Simulations