Federated Multi-Task Learning under a Mixture of Distributions