Understanding Mesa-optimization Using Toy Models