On the Unreasonable Effectiveness of Federated Averaging with Heterogeneous Data