Fix the MPI test for the CSVR thermostat in test_thermostat

[mortele]

Numbers are based on the old 2007 paper implementation and no longer correspond perfectly to the implementation following the 2008 Comput. Phys. Commun. paper (please don't ask me why). I will fix this at a later date when I hate the thermostat less.

[sigbjobo]

Please check also that the seed is independent for the different MPI tasks. This was not an issue for the serial but could be for MPI.

[mortele]

Please check also that the seed is independent for the different MPI tasks. This was not an issue for the serial but could be for MPI.

I agree that this could be an issue in general, but in this specific case the scaling is done globally with the same factor across all MPI ranks by design. I actually just generate the N(0, 1) and χ²(3N - 1) numbers on the root rank and broadcast them to the others, to make sure they are always identical.