Stopping Bayesian Optimization with Probabilistic Regret Bounds