don't save HaloField by default in high level functions

[daviesje]

Since halo fields can get quite large, and since the classes HaloFIeld and PerturbHaloField hold similar information, we really only want to cache one most of the time. So I've changed the default behaviour in run_coeval and run_lightcone to only write PerturbHaloField. This could save up to ~10% time and halve the cache requirements.

• _get_config_options is now only called on the lower-level structures, this was done to have more control over writing behaviour.
• If hooks are passed or write is set to true or false, the behaviour is the same as before
• if write==None (default) we write everything except for HaloField.

This also fixes an unrelated problem where we overpurged the ICs for the halofield since we need the velocity fields for perturb_halo_list.

In the longer term it might be a better idea to have the halo field perturbation be an in-place operation. Where we have subclasses of HaloField to indicate Lagrangian or Eulerian, which can point to the same backend arrays.

[steven-murray]

This looks good, thanks @daviesje. BTW does this fix the "writing when it's not supposed to" issues we've been having, by moving the config into the function instead of as the default?

[daviesje]

This looks good, thanks @daviesje. BTW does this fix the "writing when it's not supposed to" issues we've been having, by moving the config into the function instead of as the default?

I'm not sure, if hooks==None and write==False the first config call will set hooks == {}and pass in the default write == None since that keyword isn't in iokw. After that, since len({}) == 0, it shouldn't fill the hooks with the default write.

[daviesje]

Closing this PR as it creates issues with caching and dependence. To be revisited later