Deploying Large AI Models on Resource-Limited Devices with Split Federated Learning