The Frechet Distance of training and test distribution predicts the generalization gap