The `Scene_c3t3_item` from our 3D demos now uses an AABB tree of
cells, instead of an AABB tree of triangles. That divides the number
of primitives by 4, and save times:
Before:
```
C3t3 facets AABB tree built in 22.010006904602051 wall-clock seconds
Scene_c3t3_item_priv::computeIntersections in 0.50893402099609375 wall-clock seconds
```
After:
```
C3t3 cells AABB tree built in 13.072829008102417 wall-clock seconds
Scene_c3t3_item_priv::computeIntersections in 0.41458892822265625 wall-clock seconds
```
The gain seems mostly in the construction of the tree (40% saved), and
the gain on the queries seems lower (about 20%).
Laurent Rineau