Sample-Efficient Private Learning of Mixtures of Gaussians