Satellite specific quenching for satellites

[alexalar]

This PR adds satellite-specific quenching for satellite haloes. It does so by adding 4 new parameters for the sigmoid of frac_quench_sat, while keeping a separate 4-param sigmoid for the centrals' frac_quench_cen. There's a new parameter, upid determining if a halo is central or satellite, and frac_quench for each halo takes the frac_quench_cen or frac_quench_sat value depending on upid, using jnp.where. There's a set of duplicate files in several folders with the suffix _satfrac implementing these changes.

The plots below show results for SMDPL (no-merging). As a result of this change, there's an improvement in the fit quality of p(sSFR | Mstar), for both centrals and satellites, especially for main sequence galaxies. The improvement is not spectacular, but visible if one compares back-and-forth with the model that uses the same frac_quench for both cen and sat. For SMDPL compare to #46.

One worrisome feature is that the gradient descent gets halted because of a nan in the gradients. I suspect the covariance must be hitting some of the boundaries, perhaps because of the rho_bounds=(-0.3, 0.3), which might make the covariance non-positive definite. I want to test what happens if one uses a diagonal covariance.

[alexalar]

Here is how things look like for TNG. To be compared with #47. There is no obvious improvement in my opinion.

[alexalar]

Based on these results, it seems that results improve a bit for SMDPL, but not so clearly for TNG. This might still be useful in the future, but I feel perhaps it is worth testing first a case where the MS parameters are independent for quenched and mainseq galaxies. Potentially setting the off-diagonal covariance to zero, which would compensate the increase of free parameters.

[alexalar]

upid is missing here