Blender Git Loki

Save one memory block allocation/disposing when propagating changes from
deformed PBVH to the base mesh.

It's not real bottleneck, but crazyspace corrections looks nicer now.
Real bottleneck is normals re-calculation, which calls plenty of
fsqrt's. We could avoid this for some modifiers (which don't use
normals), but such checking would make code less controllable.