Describir: Efficient Use of GPU Memory for Large-Scale Deep Learning Model Training