AR$^2$: Adversarial Reinforcement Learning for Abstract Reasoning in Large Language Models