Forgotten Polygons: Multimodal Large Language Models are Shape-Blind