Meta-Learning Transferable Active Learning Policies by Deep Reinforcement Learning