Beyond the Federation: Topology-aware Federated Learning for Generalization to Unseen Clients