ViCToR: Improving Visual Comprehension via Token Reconstruction for Pretraining LMMs - View it on GitHub