Proper Learning, Helly Number, and an Optimal SVM Bound