When a user requests an archive download, `RepositoriesController`
attempts to extract the given ref by finding the longest match of
available branch and tag names. This is unnecessary work because unlike
other controllers, the ref name isn't ambiguous. The `id` parameter
takes the from `:ref/:filename.ext`, where `:ref` can contain any number
of slashes, and `:filename` is simply a base filename with no slashes.

We can optimize this by doing a greedy match for the reference and using
the last part as the filename. This avoid having to call out to Redis
and load large sets of data.

Relates to https://gitlab.com/gitlab-com/gl-infra/scalability/issues/124