Improve interpolation memory requirements

[adperezm]

Modify the interpolation search loop by only sending around to the rank candidates their actual points. Now we send ALL points around. This is okay for smaller point counts but not so good if one does something as mesh to mesh interpolation.

This affects not only memory requirements but execution time, as for large point counts, the local element search might not be negligible.

[adperezm]

Some testing is done in #202. Perhaps we do not need to merge everything, but some of the commits can be cherry picked.

[adperezm]

So I have tracked down some stuff. I think the main problem right now are the lists generated by the candidate rank and element searches. They can be quite large for ill posed cases. Find ways to reduce the need to use lists. Start, perhaps, changing to pure numpy arrays or so.