Fix star interactions

The idea is that we were missing some self interaction because we were setting cj_nodeID = -1 and then the condition is done in between ci and cj. I have checked the number of neighbors between swift and swift_mpi with 1 rank and they agree now.

@matthieu This is my try to fix the bug identified by Alexei. I have only one problem (not sure if it is related). When I try to write a snapshot, it happens that I receive the following error [0000] [00005.1] cell.c:cell_drift_part():3759: Attempt to drift to the past.

@aborissov can you test this please? On my test, I have the same mass evolution for both serial and parallel.