Switch to a query builder approach for the annotation counts

[marcospri]

This allows to reuse the same code for multiple cases.

Introduce the h_userids filter which with the new approach is available in all queries.

Testing

See: https://github.com/hypothesis/lms/pull/6361