Differentially Private Methods for Compositional Data