Mitigating Disparate Impact of Differential Privacy in Federated Learning through Robust Clustering