Forgotten Polygons: Multimodal Large Language Models Are Shape-Blind