Queries that do not rely on limit/offset continue to use
the existing approach, where a simple where clause is added
to the query.

Complex queries are converted into a lateral subquery.
The outer query contains the inputs.
Ideally the query would use `unnest` or a similar function
to efficiently load an array of values into a query, but
Ecto doesn't offer such functionality in a portable way.

The workaround here is to run a distinct query that selects
the column from the queryable where the value is any of the
inputs.

When associating via leaderboard, all users are associated with all scores.
It makes for a better test to only see that users scores.

This can happen when there is a through association like
A -has many-> B -belongs to-> C

When loading the A -> C association using a join, duplicate entries for C
were being returned.

The fix is to throw a distinct() on the inner query.
There may be a more elegant approach to detecting when
the distinct is necessary, but at least the results are correct now.

When A HasMany B and B BelongsTo C, then A can have the same C multiple times.
A ManyToMany association is similar.

This unfortunately required a significant duplication of the
build_preload_query code.

When the users query contains joins, the first join from the users
query to an association must target the first source in the query,
not the last one.