Shape and Time Distortion Loss for Training Deep Time Series Forecasting Models