Fix undrifted gpart in SF task

This should address the problem seen in some COLIBRE runs on HAWK.

I think the check in cell_add_spart() is wrong. The code should care about spart only. At the time where the code was written, there was no separation in the drift status of the stars and gpart. The check likely almost never fails since we usually drift a lot more gpart than anything else.

There was indeed a missing dependency for the drift_stars --> SF.

And I needed to add a piece of code to propagate the "drifted" status to empty leaves.