FedMP: Tackling Medical Feature Heterogeneity in Federated Learning from a Manifold Perspective