Improving Generalization in Federated Learning with Model-Data Mutual Information Regularization: A Posterior Inference Approach