AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation