First-Explore, then Exploit: Meta-Learning to Solve Hard Exploration-Exploitation Trade-Offs