This is an implementation that is about 1.5-2.1 times faster. It gives a result
that is on average 6° different from the old implementation. The difference is
because normals (Ng, N, N') are not selected to be coplanar, but instead
reflection R is lifted the least amount and the N' is computed as a bisector.
Differential Revision: https://developer.blender.org/D10084