Building on #6591 (review), it would be helpful if there was a scalable solution for normalizing eigenvector signs much like sklearn.svd_flip.

There is some discussion at #6591 (comment) on where it could live and how it would be applied.

This is particularly useful for library maintainers that want to compare results from different SVD implementations, though it is also necessary for users that wish to make SVD output reproducible.

• Replace the svd_flip function used in the unit tests for Add SVD support for short-and-fat arrays #6591