I agree. I've noticed that the articles like this one that use pca.components_ * np.sqrt(pca.explained_variance_) seem to cite the theory behind what PCA loadings are to explain their reasoning. On the other hand, the ones that use abs or normalize the abs of the pca.components_ seem to focus on feature importance like this.

I think this blog best explains it in its section titled "Difference Between Loadings, Correlation Coefficients and Eigenvectors."

I hope SciKit fixes this in their guides and adds a method to compute it soon.