Exploration is Harder than Prediction: Cryptographically Separating Reinforcement Learning from Supervised Learning