PipelineDatasetTypes does preserve storage classes (that's why
it's being deprecated).

First problem was just that the test assumed it could register a
dataset type after the QG was made and not run into trouble when
running that QG.

Second problem was that it was just expecting something other than what
its own code comments suggested, which was also contrary to correct
behavior.

This includes:

 - Using TaskMetadata.from_dict instead of to_dict in comparisons to
   make the intent clearer.

 - Documenting what the log-inspection utility method actually tests.

 - Making one test (test_from_pipeline_intermediates_differ)
   registering a dataset type before building a QG because that's
   what's needed in general for correctness (even though it didn't
   matter here).

 - Renaming and re-documenting another test
   (test_from_pipeline_inconsistent_dataset_types) where we were
   actually testing that building the QG and then changing a
   dataset type out from under it can be a problem.

We have a custom QG builders in the wild and they're not as
well-behaved as the main one; guard against them until we can fix them.