All in an Aggregated Image for In-Image Learning