Learning to Act from Actionless Videos through Dense Correspondences