Reasoning under Vision: Understanding Visual-Spatial Cognition in Vision-Language Models for CAPTCHA