Towards Understanding Graphical Perception in Large Multimodal Models