Added smoothing length correction limiter to Gizmo hydro_predict_extra.

For some reason, the smoothing length prediction in GIZMO sometimes goes completely wrong, causing the code to crash in space_regrid, see #271. To overcome this, I have added a limit on the increase in smoothing length during the prediction step for GIZMO. This appears to solve the issue @jwillis was having with the EAGLE_12 test (I am currently testing it on one of my machines, and it's still running after 6000 steps).

I'm not sure it is worth looking into the cause of the bad smoothing length predictions. The smoothing length update is computed based on a mixture of SPH and GIZMO variables, so this is bound to go wrong sometimes. Maybe there are smarter ways to do this.

See merge request !318

sort and drift on-demand in a consistent way. yes, even for the force loops as changes in the smoothing lengths may require a re-sort or interactions at a different level.

several changes: 1) remove init tasks and their dependencies, 2) init particles by hand in engine_init_particles, 3) make gravity tasks depend on ghost since inits are gone, 4) add hierarchical dependencies to sort tasks.