Stack More Layers Differently: High-Rank Training Through Low-Rank Updates