By specifying `key`, we get a different lazy batch loader for each
repository, which means that accessing a lazy object from one repository
will only result in that repository's objects being fetched, not those
of other repositories, saving us some unnecessary Gitaly lookups.