ViCrit: A Verifiable Reinforcement Learning Proxy Task for Visual Perception in VLMs