Make the diagonal fit the default

[scarlehoff]

It has been decided that for 4.1 we will want the diagonal covmat to be the default way in which fits are run, see this comment:
#2301 (comment)

(it would be best if this is done as one of the first PR after 4.1.0 is released, so by looking at the version of the code one can redo previous fits with or without diagonal)

In addition to make it default, we want to make sure that by using the diagonal basis we reduce the cost on memory of the fits, so that PR would be the perfect place to benchmark that and improve the performance of the fits.

• Make diagonal basis default
• Add a section specific to the diagonalization to the docs
• Add a few tests with and without diagonalization
• Benchmark the memory footprint (does it reduce? By how much?)
• Benchmark it in multireplica fits

[scarlehoff]

@jacoterh since making the diagonal fit the default will force us to redo all regression fits, please change on that PR also the pseudodata generation to not use the old names anymore #2156

[jacoterh]

Will do @scarlehoff !

[scarlehoff]

One point that I'm not sure we looked at is what to do in multireplica fits. I think you will have to move the diagonalization logic outside of the fitting_data_dict, as this function is called once per replica, so dataset_inputs_fitting_covmat should already be diagonal when you get there.

Edit:

Actually, all covmat manipulation should happen before getting to fitting_data_dict, diagonal or not.

[scarlehoff]

what is the status of this?

[jacoterh]

Unchanged sadly - I finally have time this week in Como to work on this after too many conferences and talks to prepare... I'll report here when I've made progress

[scarlehoff]

Ok! Just wanted to know at which stage were we with the "final methodology"
Make sure to have a look at #2338 before, I'm a bit worried about a growth of memory usage and resources if we are not careful here.