Treat 1-point datasets equally in sequential and parallel fits

[scarlehoff]

Due to the shape-changing nature of boolean masks we decided to just put single-point datasets in training when running in GPU. This PR removes that limitation by just accepting the point in both training and validation and setting to 0 the row and column in the inverse covmat (so the masking happens at the level of the loss)*

If this works ok, #2138 comes for free.

@RoyStegeman @achiefa I want to run a few tests first before considering this good:

• That Etr and Evl are ok for multireplica fits (they are ok for single replicas given that the regression test pass)
• Running a fit with only single-point datasets (this is actually broken in master)
• Experiments with one single dataset which is one-single-point.
• Check whether the written down replicas are really the same in sequential and parallel
  • Add a test for exactly this situation
  • Separate the .csv files per replica
  • In principle the table of written down replicas is not affected by the trick used by n3fit_data but needs to be checked

At the moment the .csv files are only stored for the last replica, once per replica and with all data in each of them. However, the information itself is correct.

(any help running the checks would be appreciated ofc, the more eyes the better)

This is needed for tree-saving reasons.

[github-actions]

Greetings from your nice fit 🤖 !
I have good news for you, I just finished my tasks:

• Fit Name: NNBOT-8653ace66-2025-02-12
• Fit Report wrt master: https://vp.nnpdf.science/SV0kXuXrTGi0eoHGE5rGtQ==
• Fit Report wrt latest stable reference: https://vp.nnpdf.science/--DkQbM1SAq7-YOePBsoiA==
• Fit Data: https://data.nnpdf.science/fits/NNBOT-8653ace66-2025-02-12.tar.gz

Check the report carefully, and please buy me a ☕ , or better, a GPU 😉!