The Asymmetric Maximum Margin Bias of Quasi-Homogeneous Neural Networks