Make the init-grav task drift the multipoles before initialising them

This bug was triggered by a rare case related to the large number of top-level cells and the absence of mesh-gravity. We could have an intermediate level between the top-level and the gravity super-level where no actual gravity operations are performed but the m-pole still needs to be drifted for the down pass. This is now done in the gravity init task.

Fixes #741 (closed).