Differentially Private Clustered Federated Learning with Privacy-Preserving Initialization and Normality-Driven Aggregation