KDTree enhancements

[apontzen]

A number of KDTree enhancements have become necessary for future projects:

• Make KDTree use npy_intp or equivalent for its indexing, which will be 64-bit on modern platforms so that larger files can be processed
• Add ability to query spheres using the KDTree for speed
• Reinstate parallel build of KDTree (it was previously found to be slow in ~2014, but unclear why this was - on modern machines seems like it works pretty well albeit not quite linearly)
• Enable KDTrees to be built in shared memory for tangos
• Move KDTree into its own module, to reflect it now won't only be used for SPH
• Switch to C++11 threading model, remove dependence on pthread

Longer-term required actions have been migrated into #770