Retention Score: Quantifying Jailbreak Risks for Vision Language Models