Deep linear networks for regression are implicitly regularized towards flat minima