User-Preference Meets Pareto-Optimality: Multi-Objective Bayesian Optimization with Local Gradient Search