Distilling Many-Shot In-Context Learning into a Cheat Sheet