Minimizing Client Drift in Federated Learning via Adaptive Bias Estimation