Distributed, communication-efficient, and differentially private estimation of KL divergence