This used to be one massive method; now it's two medium-large methods
and two very simple ones, and it's much easier to follow.

This shouldn't change the query behavior at all, but it does collapse
some UNION ALL constructs by merging their IN clauses when they are
otherwise the same - the original code had an unnecessary distinction
between CALIBRATION and other collection types (as well as some
necessary distinctions, which this preserves).

...which I just merged without a changelog entry.

This prevents query construction from failing when someone queries for
a skypix-dimension dataset type in collections where none exist.  When
the datasets do exist, we rely on the dataset table to bring in the
skypix dimension, since those don't have tables of their own, and we
need to do the same when the query is doomed to avoid triggering an
exception about skypix dimensions not being directly joinable.

This fixes an edge-case bug involving mixed CALIBRATION/non-CALIBRATION
searches introduced in the previous refactoring commit.

This involves a small internal API change for DatasetRecordStorage -
it now returns a SQLAlchemy object directly instead of one of our
SimpleQuery instances, since it sometimes now returns a UNION.

These were revealed by installing the sqlalchemy-stubs package locally.
That's something we should consider doing more broadly (e.g. in CI), as
it revealed a lot more than just these, but it seems tangled up with
migrating more fully to SQLAlchemy 2.0 APIs (there is a different stubs
package for SQLAlchemy 2), and since that's not even fully released yet
doing much more seems premature.