Dataset Distillation-based Hybrid Federated Learning on Non-IID Data