Initial dev

[jkim0731]

Functions are still in notebook cells.
It will be good enough to check how it's done, and to point out errors if any.
The results are saved in the scratch folder (also as a dataAsset), so it can be used for further analysis by duplicating the capsule or downloading it.

It does not take too long to calculate specific lambda per model - about 15 min for 34 models with 25 nested cv folds.
Surprisingly, the fit traces from mean W and those from stitched from each cross-validation (_from_splits) are very similar, and the fit is quite good. (note that even though the name says filtered_events, these are still unfiltered raw events)
It means that the models across cv folds are very similar (possibly due to stratification).
Also the good performance might be from GCaMP8s.
But these also could be due to some errors in the code.

Remaining works

• Minimizing the resultting file size (currently 9GB)
• Changing minor run params (including removing some of single models - this involves changing design matrix)
• Adjusted var_exp for binary supports
• Converting to py files