In metalearning document validation for pipeline, make sure "source.from" is also just a reference

I think this has been missed in canonical representation of pipeline document.

Moreover, it looks like the schema is also strange here. It should probably be something like {type, dataset} object or something. Currently it merges type and document reference into one object.

Also, should we have multiple types of redaction (for datasets, pipelines, problems?). Currently it was meant to be used only for datasets. So should we use dataset_or_dataset_reference there instead of just document_reference?