Bound and reclaim the runner's saved checkouts

Bound and reclaim the runner's saved checkouts

The runner already downloaded each repository once into --dir and cloned every
job from that copy, but nothing kept the copies bounded: they grew with every
tag, every branch, and every commit a job pinned by id, and a force-push left the
old history reachable through the remote-tracking reflogs. A long-lived runner
filled its disk and then could not check anything out at all.

Repack a copy that is most of the cap, dropping the objects no branch or tag can
reach and the temporary refs a killed runner leaves behind; then delete the least
recently used copy when the cap is still reached. --git-cache-size sets the cap
(10G by default, 0 keeps every copy). The repository this runner is about to
build is never the one deleted, and a copy another process holds is passed over.

Reading and writing a saved checkout now takes a per-repository lock, so a repack
or a deletion in one runner process cannot run underneath another's clone, or
underneath the site and release pushes that are written from the same cache.
The first clone skips the working-tree checkout, which nothing ever read.