Particles now get their time-step updated. Everything works. Need to improve handling of first time-step.

Updated the kick operation to the new framework. Need to do the preparation for the prediction next.

Removed the softened isothermal potential. The isothermal potential is now softened by default. Set epsilon to 0 to restore the old behaviour.

Drift on demand

Fix to the time integration bug that triggers #248. 

As written in the minutes, finding all the cells that need to be drifted can be really difficult in some convoluted situations and I'd like to avoid a full tree walk. 

My solution is as follows:
 - (Re-)Create a drift task. When a cell has active particles, it's drift task is active
 - The `runner_do_drift` function gets replaced by a `runner_do_unskip` function that does not touch the particles but just updates the status of the tasks.#
 - There is a `runner_do_drift_all` function that drifts everything for the cases where we need this (rebuild or snapshot).
 - If a pair task involves a cell that is inactive and has hence not been drifted, the pair task starts by calling the drift operation on this cell and then carries on with its regular work. Most of the drifts are done via the regular task though.

In terms of performance it looks fairly similar to the current master. But now it is physically correct.

See merge request !292

This reverts commit de558d53.

The non-MPI SodShock test starting using 0 dt steps.