cross_validate and cross_val_score raise errors_score warning but don't expose it

[amueller]

Via #11570.
Both cross_val_score and cross_validate provide deprecation warnings when an error happens during the process, telling users to set error_score. But they don't expose error_score. We either need to expose error_score in these two functions or explicitly pass error_score='raise' to _fit_and_score.

I'm pretty agnostic which of the two we do. I guess it's less likely that some of the fold will error than that some parameter settings will error in GridSearchCV.

[samronsin]

Working on this.

[amueller]

same for learning_curve. we need to check if we use _fit_and_score in other places. Or maybe we should have not changed the default in _fit_and_score but in GridSearchCV instead.

[samronsin]

Same for validation_curve as well.
There are two other occurences of _fit_and_score: in fit_grid_point and the fit method of BaseSearchCV, but errors_score is exposed in both functions.

[jnothman]

I'm happy with forcing it to 'raise'. this seems more consistent with expectation when you want to cross validate one param setting.