Towards Communication-efficient Federated Learning via Sparse and Aligned Adaptive Optimization