Stealthy Imitation: Reward-guided Environment-free Policy Stealing