This fixes the random number generator in TestGravity to be deterministic and machine-independent. It passes the quick test suite on Columbia hotfoot cluster.
With this, I was able to generate local results, then rerun new simulations and have them pass continually against my local results even when I changed the test to be bitwise identical. We probably don't need to change the actual test to bitwise identical, but I am in favor of tightening the tolerance in GravityTest__test_gravity.py. How do people feel about that?
Either way, the gold standard should be regenerated for this test.
I think it would be good to tighten the tolerance. Currently it's set to 10% of the RMS force errors. I guess now that the positions are deterministic, that could be taken down to a much lower tolerance.