For binary tasks, OvO has a shape of (n_samples, 2) which pretty much violates our standards.
I'm not sure what the best solution is apart from just breaking it as a bug-fix.
We could add a parameter and deprecate it and then remove the parameter if we really want.

The OVR classifier also has a potential issue where the decision_function for binary is (n_samples, 1) instead of (n_samples,). That also seems non-standard. I'm not sure we have other multi-output classifiers with a decision function, so I'm not entirely certain what the standard should be. All the multi-label ones do (n_samples,) according to the tests.

Found via #8022.