Learning Reconfigurable Representations for Multimodal Federated Learning with Missing Data