Differentially Private Learning of Structured Discrete Distributions