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