Hyperparameter Loss Surfaces Are Simple Near their Optima