Towards Lifelong Learning of Large Language Models: A Survey