Differentially Private Algorithms for Learning Mixtures of Separated Gaussians