Distribution-Aware Federated Learning for Diabetes Prediction Using Tabular Clinical DataUnder Non-IID and Class-Imbalanced Settings