Reconstructing and Querying ML Pipeline Intermediates