Instance Selection for Dynamic Algorithm Configuration with Reinforcement Learning: Improving Generalization