The Preloader relies on other objects to bind the retrieved records to their
parents. When executed across a hash, it assumes that the results of
`preloaded_records` is the appropriate set of records to pass in to the next
layer.

Filtering based on the reflection properties in `preloaded_records` allows us to
avoid excessive preloading in the instance where we are loading across a
`has_one` association distinguished by an order (e.g. "last comment" or
similar), by dropping these records before they are returned to the
Preloader. In this situation, we avoid potentially very long key lists in
generated queries and the consequential AR object instantiations.

This is mostly relevant if the underlying linked set has relatively many
records, because this is effectively a multiplier on the number of records
returned on the far side of the preload. Unfortunately, avoiding the
over-retrieval of the `has_one` association seems to require substantial changes
to the preloader design, and probably adaptor-specific logic -- it is a
top-by-group problem.