Instead of making each ClipScrollGroup per stacking context and
ClipScrollInfo combination, make them per ClipScrollInfo. This should
reduce the amount of work done per stacking context and is the first
step toward accepting all coordinates relative to reference frames.