Fixed a moronic bug. Unskipping MPI sends and recvs for gravity should not be done *only* for local pairs. Obviously....

Removed unnecessary dependency between the out-going gravity ghost and the self-grav and pair-grav tasks.

Open only once dependency file and add an option in debugging-checks mode for writing task dependency graph

Add information to stdout and timestep.txt about what happened in the step (rebuild, repartition, snapshot, ...)

An attempt to tune these calls without affecting the other parts which work better with buffered MPI

Implement simple thread barriers to be used on systems where the POSIX implementation does not provide them. We default to the POSIX ones when detected at configure time.

Use the threadpool to collect data over the top-level cells at the end of the timestep. Also only loop over the cells that have tasks in their hierarchy.

Instead of listing local cells only, list cells that have a tasks associated with them on the local node.

All gpart now find all of their interactions following correct boundary condition check in runner_do_grav_long_range()

Can now run through all the tasks but still missing some interactions coming from the close-by pairs over MPI

All-reduce all the top-level multipoles after a rebuild() to make sure we have all information in hands to construct the tasks.

Fixed the gravity bugs reported in #363. Need to find a good rebuild criterion for gravity-only runs.