Implemented a simple heuristic that should keep the number of cached
features per layer reasonable - and thus lower the amount of consumed
memory and CPU for big layers.
Instead of calculating feature count (which may add extra overhead),
let's directly try to build index for layers. Only when a certain limit
of features is reached (currently 1 million), the indexing is cancelled
and layer is marked as non-indexable.
The maintenance of three R-trees was too complicated, using a lot of memory
and the performance benefit was questionable :)
The approach with just one R-tree leads to much simpler code
Removed k-NN queries in the process - they were not used in QGIS code. They can be
reintroduced later, but there is not much use for them right now...
