This is a quick fix for the aborts reported in #1129 by moving the call to disable the GSL error handling from ccl_parameters_create to ccl_cosmology_create.

This uncovered a bunch of other issues though. From the python side, ccl_parameters_create isn't being used any more.
It is still used from within the C code though, mainly in halofit and the mu-Sigma MG stuff. So now there are two places in the code that deal with the non-trivial aspects of the ccl_parameters struct, namely the allocation of the massive neutrino arrays and MG arrays.

Speaking of these arrays, right now you can do some pretty nasty stuff by accessing arbitrary memory because N_nu_mass can be changed from python. I added a fix for one memory leak but there might be others.

import pyccl as ccl
import numpy as np

cosmo = ccl.CosmologyVanillaLCDM()

print(f"{cosmo._params.N_nu_mass=}")
print(f"{cosmo._params.m_nu=}")

print("Set cosmo._params.m_nu")
cosmo._params.m_nu = np.array([0.1, 0.2, 0.3])

print(f"{cosmo._params.N_nu_mass=}")
print(f"{cosmo._params.m_nu=}")

print("Set cosmo._params.N_nu_mass")
cosmo._params.N_nu_mass = 10
print(f"{cosmo._params.N_nu_mass=}")
print(f"{cosmo._params.m_nu=}")

gives

cosmo._params.N_nu_mass=0
cosmo._params.m_nu=[]
Set cosmo._params.m_nu
cosmo._params.N_nu_mass=3
cosmo._params.m_nu=[0.1, 0.2, 0.3]
Set cosmo._params.N_nu_mass
cosmo._params.N_nu_mass=10
cosmo._params.m_nu=[0.1, 0.2, 0.3, 0.0, 5e-324, 1.6e-322, 2.1470999305e-314, 1.5e-323, 9e-323, 1.5543504411933145e-303]

I'm aware that _params is supposed to be private but it shouldn't be that easy to mess with C memory.