[MRG+1] Fix cross_val_predict behavior for binary classification in...

[MRG+1] Fix cross_val_predict behavior for binary classification in decision_function (Fixes #9589) (#9593)

* fix cross_val_predict for binary classification in decision_function

* Add unit tests

* Add unit tests

* Add unit tests

* better fix

* fix conflict

* fix broken

* only calculate n_classes if one of 'decision_function', 'predict_proba', 'predict_log_proba'

* add test for SVC ovo in cross_val_predict

* flake8 fix

* fix case of ovo and imbalanced folds for binary classification

* change assert_raises to assert_raise_message for ovo case

* fix flake8 linetoo long

* add comments and clearer tests

* improve comments and error message for OvO

* fix .format error with L

* use assert_raises_regex for better error message

* raise error in decision_function special cases. change predict_log_proba missing classes to minimum numpy value

* fix broken tests due to special cases of decision_function

* add modified test for decision_function behavior that does not trigger edge cases

* fix typos

* fix typos

* escape regex .

* escape regex .

* address comments. one unaddressed comment

* simplify code

* flake

* wrong classes range

* address comments. adjust error message

* add warning

* change warning to runtimewarning

* add test for the warning

* Use assert_warns_message rather than assert_warns

Other minor fixes

* Note on class-absent replacement values

* Improve error message