FigStep: Jailbreaking Large Vision-language Models via Typographic Visual Prompts