Rethinking Post-Unlearning Behavior of Large Vision-Language Models