Collaborative Learning of Discrete Distributions under Heterogeneity and Communication Constraints