FedOC: Optimizing Global Prototypes with Orthogonality Constraints for Enhancing Embeddings Separation in Heterogeneous Federated Learning