Federated Learning from Vision-Language Foundation Models: Theoretical Analysis and Method