Fixed the random generation of the GMFs

[micheles]

See #9073.
For an historical accident the random generator was reset for each gsim, now instead there is a single random generator for all gsims. This is consistent with how it is done in the ConditionedGmfComputer and will allow me to remove some duplication between the two computers.

[J-Cohen-Gabor]

The bug appears to still be present in the 3.18 release (I generated a similar plot breaking down the random values by GMPE as in the original issue). While the random seed has been broken out by GMPE, the compute_all function loops through the rupture which is actually what links the seed. A structural fix would require changing the seed in compute_all to something entirely different - a simple fix is to simply change the gmf seed by 1

def compute_all(self, mean_stds, max_iml=None,
                    cmon=Monitor(), umon=Monitor()):
        """
        :returns: DataFrame with fields eid, rlz, sid, gmv_X, ...
        """
        self.init_eid_rlz_sig_eps()
        rng = numpy.random.default_rng(self.seed + 1) #simply add 1 to the code here fixes the bug