On the Power of Pre-training for Generalization in RL: Provable Benefits and Hardness