Evaluate stateful LoD streaming

[Ralith]

The current design computes the desired set of tiles to load from scratch each frame. This is fast enough for most purposes, but there might be room for improvement:

On each update, for each resident chunk, evaluate whether it should be split or merged according to the LoD predicate. When a split or merge occurs, recursively evaluate the newly introduced chunks.

This avoids traversal of the quadtrees and perhaps hash-table lookups for coherent updates. Performance might be worse for incoherent updates, but those are of less interest.