Analysis of Thompson Sampling for Gaussian Process Optimization in the Bandit Setting