Training for the Future: A Simple Gradient Interpolation Loss to Generalize Along Time