Hierarchical Reinforcement Learning for Zero-shot Generalization with Subtask Dependencies