Training data generation with CLASS

[HTJense]

Work in progress for a CLASS interface. Will merge this when it is close enough to parity with the CAMB interface.

Thanks to Alex Nemtinova for initiating this and aiding me in fixing bugs relating to my ignorance when it comes to CLASS.

To match CAMB parity, we will need auto-generation of:

• derived parameter handling
• CMB power spectra (both lensed and unlensed)
• CMB lensing spectra
• linear matter power spectra
• non-linear matter power spectra
• H(z) (and other background) evolution
• sigma8(z) evolution
• angular diameter distances

On top of that, it might be nice to:

• automatically infer lmax for cmb spectra
• automatically infer the necessity of unlensed/lensing/lensed power spectra computation

The automatic setting inference is currently done via the "auto: True" flag in the extra_args block in the yaml. It's probably better to have everything done by the end user however, as class has too many accuracy settings that could realistically be inferred by the code (at least for camb, the only real parameter that has to be inferred is the redshift grid for quantities that vary with redshift, like sigma8(z)). It might be easier to parse the parameters and halt if there's an inconsistency (like lmax_calculated < lmax_requested), rather than try to parse the requested parameters and modifying the settings accordingly...

All work should be limited to writing the spectra_classy.py file, as the remaining interfacing is handled automatically by the pipeline from there.

[HTJense]

In principle most of the things we would like to see from class should be present. I tried generating a few (O(10^3) samples) minimal NNs to check that the behaviour is correct, and I think it matches most of the CAMB interface now.