Compressing Many-Shots in In-Context Learning