Leveraging Foundation Models for Multi-modal Federated Learning with Incomplete Modality