Prototype-based Heterogeneous Federated Learning for Blade Icing Detection in Wind Turbines with Class Imbalanced Data