Learning MDPs from Features: Predict-Then-Optimize for Sequential Decision Problems by Reinforcement Learning