Small change to derive_predictors_and_scores for speed + normalization

[PalkaPuri]

I noticed that generating a combined DF of predictors and scores was taking very long for large datasets (my kernel kept dying). This operation was very compute expensive because we were using df.merge which involves searching through column values to find matching rows. However, we can get away with using something simpler like pd.concat since all predictor/score DataFrames inherit the grid from local_windows and hence the rows match by design. This should help clear the speed bottleneck.

I also updated the function to return eCDF normalized values as an option

[PalkaPuri]

Added option to scale values based on empirical CDF
(rescaling using min and max, which we were implementing before, does nothing as the predictors are already scaled to be between 0 and 1)

[rfl-urbaniak]

Please pull origin from the current version of staging, make sure you resolve all conflicts and pass all the tests.

[PalkaPuri]

@rfl-urbaniak @dimkab Just finished going over this branch. The following changes were implemented:

1. final derivedDF is now generated by concatenating each predictor/score DF on index instead of using df.merge which was more computationally intensive
2. Added time logging for generation of derivedDF and local_windows
3. Updated the UserWarning in case of NaNs : previously we said X frames are dropped from derivedDF. But that is incorrect as the number X corresponds to rows of the df (which may be coming from any number of frames of the data).
4. Updated add_scaled_scores option to scale values according to the empirical CDF. Previously scaling using min/max did not do anything as all predictors/scores are already scaled to be between 0 and 1.
5. The output of the corresponding test notebook was changed due to these updates

Note: After merging these changes we would need to run all the doc notebooks again, as 2,3, and 4 can potentially change the outputs of the cells. I did not do this just yet so as to not overwhelm you both with 100 file changes in one PR.

[rfl-urbaniak]

I am not convinced that replacing merge with concat has no strange consequences at least for some of the predictors. As a sanity check I re-run communicators inference with your proposed modification and the posterior marginals are significantly different from the ones we currently have. I think this modification, if indeed correct, needs additional explicit tests that ensure proper functioning, involving predictors other than velocity too.