GenRecal: Generation after Recalibration from Large to Small Vision-Language Models