Õptimal Differentially Private Learning of Thresholds and Quasi-Concave Optimization