Free sort indices and task arrays before repartitioning

Freeing the sort indices and tasks arrays before repartitioning makes that memory available so 
that we have more room for accepting particles from other nodes.

Part of #98.

Also adds a useful debug function that reports the current memory
usage of the process tree (i.e. call from main thread to get all usage).
Same numbers you get out of the top program.

Free the task arrays and cells are refactored into simple functions to cut down
on cut and paste.

See merge request !392

Also updated the Pressure-Entropy SPH scheme to use the new definition of the number of neighbours and its derivative. Back to correct results.

Lots of refactoring to tidyup code reuse

Updated version number to 0.6.0 to reflect the large number of changes to the threadpool, drifts, MPI and scaling performance in the last few months.

Periodic gravity speed and accuracy improvements



See merge request !393

Threadpoolize all the things

These are all the remaining function from the list in #338. I've tested this as far as possible on my laptop, i.e. MPI and gravity *should* work.

See merge request !384

Merge branch 'threadpoolize_all_the_things' of gitlab.cosma.dur.ac.uk:swift/swiftsim into threadpoolize_all_the_things

convert engine_make_self_gravity_tasks to use a threadpool_map. still have to create the grav_ghost tasks sequentially, but i think that's ok since there are 128 times fewer of them.

Engine barrier

Following the results in the `threadpool_task_plots` branch, I've replaced the elaborate hand-crafted `engine_barrier` function by two `pthread_barrier`s.

As a result, the runner threads should all start, and synchronize, faster.

See merge request !386

engine_launch is only ever called with all threads, remove the option to limit the number of running threads.

engine_launch is only ever called with all threads, remove the option to limit the number of running threads.

Keeps the memory use down

Added the EAGLE_6 test case. This is a very small cosmological volume not representative of the workload of large runs.

Allocate a separate sort array for each dimensions.

As discussed, fixes #334.

See merge request !380