Provable Guarantees for Gradient-Based Meta-Learning