Differentially Private Federated Learning With Time-Adaptive Privacy Spending