📐 DeepSketcher: Internalizing Visual Manipulation for Multimodal Reasoning