Unfortunately, our tag system for allowing nans does not work with pipelines. Lets say we have a pipeline with two steps and the final step does not accept nans:

1. If the first step is an Imputer, then the pipeline accept nans. For example: make_pipeline(SimpleImputer(), LogisticRegression())
2. If the first step is a StandardScalar (which accept nans and leaves them along), then the pipeline does not accept nans. For example: make_pipeline(StandardScalar(), LogisticRegression())

We likely need a "output_nan" tag to reliability give a pipeline a "allow_nan" tag.