Bayesian Optimization with Noise-Free Observations: Improved Regret Bounds via Random Exploration