Loosened the model-free optimisation tests as different numpy versions were affecting the precision.
The checks of the model-free params and chi2 use self.assertAlmostEqual() again, as numpy was
causing float differences at around the machine precision mark.
