Bayesian Active Learning for Optimization and Uncertainty Quantification with Applications to Protein Docking