Improve geom_pointgrid performance for very large input data sets and grids with a distance threshold

[nevrome]

Performance and memory-footprint can be significantly reduced by filtering out all point-grid distances above a certain threshold. I think this threshold can not be easily determined automatically, but it could be a user-provided input to the core algorithm (with a default of NULL).

However filtering the distance table may produce a state that the algorithm can not solve any more. I don't know if it would fail then with an error or even end up in an endless loop. Adding this feature involves figuring out when to abort gracefully in case of failure.