SO Foreground marginalization

[HTJense]

This is a work-in-progress for the foreground marginalization & fg-marginalized likelihood codes for Simons Observatory. The methodology is comparable to the original pliklite code for Planck (gibbs sampling of the CMB bandpowers).

Current steps that are to be done by the end of this PR:

• Create a working script & internal code mechanics for the CMB extraction from SO data
• Setup the initial preset parameters (e.g. parameter ranges, restrictions, extraction ranges, etc.) for SO.
• Utilities for plotting the results & compressing the result into a new cmb-only sacc file.
• A CMB-only likelihood for the cmb-only data.

Steps that might be needed but can be done at a later stage:

• The full systematics modeling (the current implementation only allows for calibration & bandpass integration uncertainties, but we might have alm rotation uncertainties or beam leakage, which aren't included yet).

PR is mainly for tracking progress updates.
Will request a review once the branch is in a state where it can be merged.

[codecov-commenter]

Codecov Report

Merging #158 (0fb9a5a) into master (1b7c793) will decrease coverage by 4.74%.
The diff coverage is 19.79%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #158      +/-   ##
==========================================
- Coverage   74.61%   69.87%   -4.74%     
==========================================
  Files          33       36       +3     
  Lines        2029     2221     +192     
==========================================
+ Hits         1514     1552      +38     
- Misses        515      669     +154     

Files	Coverage Δ
soliket/__init__.py	100.00% <100.00%> (ø)
soliket/cmbonly/__init__.py	100.00% <100.00%> (ø)
soliket/mflike/__init__.py	100.00% <100.00%> (ø)
soliket/cmbonly/cmbonly.py	20.48% <20.48%> (ø)
soliket/mflike/ForegroundMarginalizer.py	16.98% <16.98%> (ø)

[itrharrison]

This looks generally good! I think that I would be happy to merge after you add a test.

Ideally the test would check against both a fixed likelihood value, and also verify that the CMBOnly likelihood is consistent with what one would expect from the MFLike likelihood (i.e. the validation you have done elsewhere). But just against a fixed value would do if that isn't easy.

In the long run I am also not convinced that the foregroundmarginalizer stuff belongs in SOLikeT. It seems like it would be better as its own package, with SOLikeT as a requirement. Happy to do that some time in the future though.

[itrharrison]

Would CMBOnly be the correct camel case? :-p

[cmbant]

I would have thought the code for doing the marginalization would be better a separate package, and just include the public-facing pre-marginalized likelihood in SOLikeT? (or it could go in mflike)

Not sure why you have np.exp(rat) in the acceptance criterion (can't use default Cobaya sampling for some reason?).