One Stone, Three Birds: Prototype-Enhanced Federated Learning for Mitigating Data Scarcity, Imbalance, and Heterogeneity in Blade Icing Detection Across Distributed Wind Farms