Denevil: Towards Deciphering and Navigating the Ethical Values of Large Language Models via Instruction Learning