Towards a Pretrained Model for Restless Bandits via Multi-arm Generalization