Implement CHARM as a bias model

[maho3]

This PR implements CHARM as a halo biasing model. Things done:

• Addition of cmass/bias/charm subfolder, including pretrained weights and code to load and apply them
• Addition of code in cmass.bias.rho_to_halo to optionally use and apply CHARM
• Since CHARM is trained at fixed N=128 resolution, rho_to_halo includes code to split larger volumes into batches and process them serially
• CHARM requires the density field at z=99. cmass.nbody simulators have been retro-fitted with code to save snapshots at these redshifts.
• Additional configuration arguments in cmass/conf/bias to specify which bias model to use
• Addition of colossus cosmologies to scale initial density fields with growth factors

Still todo:

• [ ] Cleanup integrate_ltu_cmass unnecessary imports, comments, and printouts
• [ ] Document helper functions in integrate_ltu_cmass

[maho3]

Note, BORG simulators are slightly different than FastPM simulators.

The raw transfer function application results in more small scale power in BORG than in FastPM:

However, the PM outputs result in more FastPM power than BORG:

This results in a ~5% bias in the CHARM halo field.

This may be later corrected by switching to FastPM, or retraining CHARM.