We stopped using these code paths outside of tests on DM-24288, and
removing them is a huge simplification for Registry.

In detail:

 - DatasetRef.components has been removed, along with the flatten and
 allRefs methods.  This means DatasetRefs now only have two states
 (resolved and unresolved), instead of it being unclear whether an
 empty dict means there are no components or that we don't know about
 the components.  DatasetRef.makeComponentRef has been added to make it
 easy to make a ref for a component dataset from the parent ref.

 - All "recursive" kwargs have been removed, along with some
 occasionally-complex logic needed to implement them.

 - The dataset_composition table has been removed.

It is worth noting that I found a couple of spots where we were still
calling insertDatasets with recursive=True (ingest and import from
yaml).  So I think we were still generating unnecessary database
rows sometimes, including in contexts where a lot of datasets were
involved.  I don't think those rows were ever used after insertion,
which is why this didn't break anything.

We now just implement __repr__ and let __str__ delegate to that, and
aim to make it informative and non-confusing because we've never been
able to make the string eval'able as in the ideal case.

Note that the biggest problem with the old repr (which is what is used
by __str__ in built-in containers of data IDs, regardless of what we'd
prefer) is that it printed the full DimensionGraph, which often has
implied dimensions and hence more "keys" than the mapping interface
actually provided access to.  That it impossible to infer what the
dict-like form actually was, and it's the dict-like form that people
actually want to know.

We now have an option to control whether components are included in
the results.  We expect users to consider them noise if the parent is
also included in the results, so the default is now to not include them
unless the parent is not in the results.