Broke one of my own rules by doing too much in a single commit: implemented runner_extra_ghost, moved gradient calculations for both gradient types to new gradient loop, fixed bug in engine.c involving the number of links (which is larger with an extra loop), streamlined gradient methods, documented most GIZMO specific files. Second loop works.

Now at least as fast as tkask version

Same behaviour as old task based drift

Add entropy to the energy diagnostic file. Use an MPI_Reduce instead of MPI_Allreduce to get the values

add a third communication step which exchanges the particles after the kick, just to make sure we have the correct ti_end_min values for each foreign cell.