Separate MS params and Satellite specific quenching

[alexalar]

This PR combines the model updates from #49 and #48, adding separate main sequence parameters for mseq and qseq galaxies, and also adding an independent frac_quench sigmoid function for centrals and satellites.

[alexalar]

Here are the results for SMDPL. To be compared with:

• Default model: Remodel with smoothly_clipped_line #46
• Separate MS update: Separate main sequence params for qseq and mseq galaxies #49
• Satellite specific quenching: Satellite specific quenching for satellites #48

This is the best version so far, and for the first time we get good sSFR distribution for low mass galaxies at z=0, for both centrals and satellites. The loss function value is also clearly the lowest (20% better) than previous best results.

[alexalar]

Similarly, here are the results for TNG. To be compared with:

• Default model: Fitting tng #47
• Separate MS update: Separate main sequence params for qseq and mseq galaxies #49
• Satellite specific quenching: Satellite specific quenching for satellites #48

The improvements in this case are not so significant. The sSFR of satellites for low mass galaxies at z=0 also get better, but the other differences remain.

[aphearin]

@alexalar I fixed some bugs in the last few commits. The reason the current PR checks fail is because the test_gradients.py test module fails on account of the parameter u_lgmu_lo_mh_hi having zero-valued gradients. It is possible that the issue is just the way the unit test is written, or it is possible that the issue is with the model formulation, I'm not sure. One way to check would be to check for non-vanishing gradients in the real loss function you are using in your optimizations, and not in this simple toy problem. If you find that theu_lgmu_lo_mh_hi parameter has non-zero gradients when applied to a real problem, then that means this current test failure is harmless, in which case we might should update the gradient test.