To Generate or to Retrieve? On the Effectiveness of Artificial Contexts for Medical Open-Domain Question Answering.