A Survey on the Honesty of Large Language Models