Performs correctly in this scenario: https://vimeo.com/234564499

Remaining considerations:

1. This is a first approximation to an appropriate "definition of UP." Currently an axis of the object is constrained to a projection of its initial direction, the projection is orthogonal to the target normal. There may be a more intuitive definition in a later version. This implementation, translated from linear algebra, should be easily adaptable to a future version of UP.
2. Disappointingly, there was not an easy way to access the target normal for the shrink type of 'nearest vertex'. So as to not disrupt existing code, I have left the align-to-normal transformation disabled in that case.
3. Surface normal interpolation (smoothing) has not yet been implemented.

Fixes T47440