Optimizes Ci::Pipeline#latest_successful_for_refs to load just the
needed branches in two queries. Instead of loading all branches in one
query then filtering the needed branches in memory.

Changelog: fixed