ViCor: Bridging Visual Understanding and Commonsense Reasoning with Large Language Models