Follow up to #16112

EDIT: the tag is now requires_y

This PR adds a is_supervised tag and uses it to raise an error in _validate_data() if:

• a supervised estimator is passed y=None
• a non-supervised estimator is passed y!=None

This PR is mostly motivated by the fact that since #16112, supervised estimators that are passed y=None will fail with a tuple unpacking error instead of a nicer error message.

This comes with a bunch of complications:

• This forces supervised estimators to call _validate_data(X, y), instead of validating X and y separately. Since calling check_array(X, ...) and then check_array(y, ...) isn't in general equivalent to calling check_X_y(X, y, ...), I had to introduce a way to check X and y separately when calling _validate_data(X, y, ...). This is really ugly. It will also definitely not work for third-party estimators that inherit from ours.

• Some estimators like OneClassSVM, IsolationForest, and RandomTreesEmbedding are unsupervised, but they generate a random y that is passed down to their base classes and validated there with _validate_data(X, y). This makes _validate_data fail because of the check mentioned above. So we need custom checks in the validation of the base classes in each of these. This is really, really ugly.

• Since _validate_data assumes the tag exists, this means that all estimators using validate_data must also support the tag. IMO, that's fine.