On Cross-Layer Alignment for Model Fusion of Heterogeneous Neural Networks