How should we handle non-feature / non-target columns for PipeOps that generate new samples

[mb706]

E.g. supersampling, smote etc. -- currently the best we do is throw an error? The more widely used things like weights should probably be supported.