FedCoLLM: A Parameter-Efficient Federated Co-tuning Framework for Large and Small Language Models