Add support for non-linear models

[tilmantroester]

This is a draft implementation to address #174 based on discussions with @jablazek, @paulrogozenski, @chrgeorgiou. It currently supports perturbative IA and galaxy bias models but it should be extendable to halo-model-based models. An example (examples/des_y1_3x2pt/des_y1_3x2pt_PT.py) showcases this functionality and compares the results to a direct CCL calculation of the same quantities. They seem to match well!

Allowing support for non-linear quantities requires extensive changes to how firecrown handles two-point statistics and passes around cosmology objects. The main changes are

• Add a CosmologyBundle class that bundles the ccl.Cosmology object with associated data, such as PT calculators. This class gets passed around as cosmo instead of the ccl.Cosmology object downstream of the top-level Likelihood call.
• Add a TracerBundle class that bundles a ccl.Tracer object with information about the associated 3D field and ccl.nl_pt.PTTracers, where relevant.
• Add a CosmoSystematicsLikelihood likelihood class that can account for cosmology-level systematics. The concrete example here is the new class PTSystematic, which populates the PT workspaces and computes the basic perturbative quantities that are later used in TwoPoint.compute
• Changed how Source.create_tracer and TwoPoint.compute work to deal with the multiple tracers that are now required per Source
• Add a TattAlignmentSystematic and PTNonLinearBiasSystematic systematic that make use of all this new functionality
• Some refactoring in number_counts.py to more generally deal with magnification bias

A plot showing the agreement with CCL: This PR currently uses the pt_enh branch of CCL, since it required some extra functionality there.

[tilmantroester]

@marcpaterno, @vitenti Do you have comments or suggestions on the overall structure of the implementation? There's still quite some clean-up to do but I want to make sure you're on board with the design before investing more time in writing and changing the code.

[vitenti]

@tilmantroester We are still reviewing the code structure and figuring out how to better integrate with design we were developing for the toolset.

[marcpaterno]

@tilmantroester We have just cut firecrown release v1.3.0, containing (among other things) the new version of the likelihood "factory script".

Would you like this code to make it into a v1.4.0, which we could cut in the next week or two? As of our last meeting, our plan was to release this code to be used in the forecasting project (DESC project 291) and to make the generalization/refactoring we discussed as a task to be done after that release.

[tilmantroester]

I'm busy until next week but after that I'll try to get this ready to be reviewed. I'm hoping to get this done before the holidays, so that we have the models ready for the forecasting project.

[tilmantroester]

This should now be ready to be looked at again. I've added a unit test which automatises the example scripts. This should now also handle combinations of PT and non-PT tracers, following a similar approach to @paulrogozenski.

One thing to discuss is whether to use the firecrown.likelihood.Cosmology everywhere. Right now the methods of Likelihood still take pyccl.Cosmology arguments, and only for certain downstream cases Statistic.compute, the firecrown Cosmology class is used.

[tilmantroester]

This is now finally passing all tests.