The chunked stuff is no longer needed for the ground tiles.
... I (Mathias Froehlich), believe that it is more performant (for the GPU) to
build the biggest indexed sets possible at the expense of tight culling.
I (Mathias Froehlich) also know that the CPU is more performant on smaller chunks
building the old flat groundcache on such tiles. :)
So this code *was* to balance these two contrary requirements to some degree.
Now we have optimized datastructures for drawing and differently optimized
datastructures for intersection tests.