Optimal Rates for Regularization of Statistical Inverse Learning Problems