White-box Multimodal Jailbreaks Against Large Vision-Language Models