On Fast Leverage Score Sampling and Optimal Learning