Perspective-Aware Reasoning in Vision-Language Models via Mental Imagery Simulation