FedMAX: Mitigating Activation Divergence for Accurate and Communication-Efficient Federated Learning