If pipeline output is not a standard DataFrame predictions, pipeline run generation fails and terminates runtime

It should not.

For example, if you have a pipeline which embeds DataFrame as cells. This is not JSON serializable and it fails (not sure why is not JSON serializable though using our generator, but that is what currently happens).

We should just not generate predictions in such case. Maybe pipeline run interface should get as an argument is_standard_pipeline from runtime and not try recording predictions in such case.