When transforming a geometry, we often apply the transposed inverse
to normals / custom normal data. However, that matrix can still contain
scale from the original matrix. That scale has to be removed so we can
avoid also scaling the normals.
I used the opportunity to remove the duplication between mesh and curves
processing of the custom normals, and to formalize an optimization to
skip the final normalization of each vector if the transform is such
that it isn't necessary. The new functions don't fit beautifully into
their public headers, but I don't know of a better place for them.
Pull Request: https://projects.blender.org/blender/blender/pulls/142896
3936d7a93e