IDEATOR: Jailbreaking Large Vision-Language Models Using Themselves