Differentially Private Release and Learning of Threshold Functions