Multimodal Federated Learning with Missing Modality via Prototype Mask and Contrast