Remove T_CMB and T_ncdm as constants (reloaded)

[nikfilippas]

This one completely decouples T_CMB and T_ncdm from the physical constants.
It also exploits the CCLParameters framework that was there to make the C-level cosmological parameters communicate with Python, and port their setting-up to Python fully.

[damonge]

Looking good.

[damonge]

Can we also:

• Introduce T_ncdm as an initialisation parameter of Cosmology.
• Make sure the C-level versions of T_CMB and TNCDM are not used at all, even at the C level, and that instead the cosmology parameters are used.

[nikfilippas]

Replying to both your general and your specific comments:

• T_ncdm is already an initialization parameter of Cosmology. It doesn't show in the diffs because this PR is based on nu_Tncdm_v3 where I first implemented it.
• Luckily the C-level functions that use the Ts all do it through the cosmology object.
• I went through the entire C library to see if there are any rogue uses of these parameters, but I think I have cleared them up. To verify, I set them both to 1e+300 at the C level, and re-ran tests and benchmarks, and they ran fine.

[damonge]

OK, sounds good. See my comment about deprecation message for T_CMB and renaming the new OmNuh2 function

[nikfilippas]

@damonge all done + neutrinos (kind of) debugged!
The fix is:

>>> ccl.CosmologyVanillaLCDM(m_nu=0.1, m_nu_type="single")["m_nu"]

array([0.1, 0. , 0. ])  # old result
[0.1]  # new result

So now, len(m_nu) matches the number of massive neutrinos, which is what this parameter originally intended.

[damonge]

Can we start doing this properly with omega_x?

[nikfilippas]

Yea this is already tested in test_background.py. In v3, we're going to delete those tests, or I can delete them now if you prefer.

[damonge]

OK, good. Can we add just one test, for my own peace of mind, that verifies that we get the same answer with omega_x and with Omeganuh2?

[nikfilippas]

They both call the same C function, ccl_Omeganuh2:

double ccl_omega_x(ccl_cosmology * cosmo, double a, ccl_species_x_label label, int *status)
{
  // If massive neutrinos are present, compute the phase-space integral and
  // get OmegaNuh2. If not, set OmegaNuh2 to zero.
  double OmNuh2;
  if ((cosmo->params.N_nu_mass) > 0.0001) {
    // Call the massive neutrino density function just once at this redshift.
    OmNuh2 = ccl_Omeganuh2(a, cosmo->params.N_nu_mass, cosmo->params.m_nu,
                           cosmo->params.T_CMB, cosmo->params.T_ncdm, status);
  } else {
    OmNuh2 = 0.;
  }

but there is a slight difference. ccl_omega_x returns zero if N_nu_mass <= 1e-4 whereas ccl_Omeganuh2 returns zero if N_nu_mass is exactly zero. I have removed the check from ccl_omega_x in order to let ccl_Omeganuh2 do the job, and now both calls should be identical, so I don't feel there is need of unit testing, since the input of ccl_omega_x is directly piped into ccl_Omeganuh2.