A Riemannian gradient flow perspective on learning deep linear neural networks