Gaussian Process Optimization in the Bandit Setting: No Regret and Experimental Design