Learned Regularization for Inverse Problems: Insights from a Spectral Model