Discovering Collaboration from Novelty: Random Network Distillation for Clustered Federated Learning