Upgrade scikit-learn to 0.24.1

[desilinguist]

See full changelog. Since this has many API changes, this might be a bit of work.

[desilinguist]

Going through the changelog, it looks like the deprecation of coef_ and intercept_ in naive bayes estimators might be an issue. See PR here.

[desilinguist]

We should also check the default parameter values for the various learners and make sure we are using the recommended values for our parameter grids and defaults.

[desilinguist]

Here's a running list of all the changes this will entail:

• [x] Setting the new keyword argument error_score to "raises" in the GridSearchCV() call made in Learner.train() since we want to raise an exception if there was any problem with fitting the estimator. This is necessary because the new scikit-learn default is to simply return a nan as the fit score in case of a problem.
• [x] LinearRegression models in scikit-learn now support a new keyword argument positive which can be set to True to use Non-negative Least Squares (NNLS) regression. This is probably something we want to enable in SKLL since it could be useful in RSMTool. Note that this will require a fix for #656 since it requires non-sparse input.