Decrease cost of halo regions by pruning with a kdtree search

[kuberry]

Currently, a halo size of distance h is specified for MPI jobs and all neighbors on adjacent processors inform nearby processors if they contain neighbors within h of the boundary.

After building halo in the traditional way, and then performing neighbor search with knowledge of adjacent off-processor neighbors, this neighbor list can be used to prune the halo region, resulting in smaller column maps and less off-processor communication.

Pruning of neighbors lists based on some criteria can be a separate PR.