Efficient Policy Learning from Surrogate-Loss Classification Reductions