Consider that sprite rendering, at least from the cpu end, is a similar problem space to physics + dynamictrees:

Ideally:

• Each drawdepth is on its own dynamictree. We have far fewer of those than collision layers so we would probably only have 20 or 30 and not 100.
• When we get sprites to draw we can sort per dynamictree and not globally which is much faster and means we can dispatch sprites for drawing much faster.
• Cleaner separation when we want to do overlays in-between drawdepths.
• Per-layer drawdepths is useful in some rare cases where you really don't want 2 entities.