Gradient-trained Weights in Wide Neural Networks Align Layerwise to Error-scaled Input Correlations