The Git access checks need access to objects which have been newly
introduced for a set of given changes. These new objects used to be
retrieved via a revision walk, where preexisting references have been
excluded and new references included. The resulting set of objects is
then the set of objects which haven't previously been referenced by
anything else. Depending on the number of references, walking all
preexisting ones can be quite expensive and easily take dozens of
seconds.

In the past, we have thus converted our LFS pointer checks to use the
quarantine directory to list new LFS pointers. The quarantine directory
hosts all objects which have been uploaded by a git-push(1), and can
thus directly tell us about which objects are new and which aren't. This
has sped up LFS pointer checks in case we have a quarantine directory by
3 orders of magnitude in biggish repos with many refs, going from up to
20 seconds down to 20 milliseconds.

Given that commits can be front-loaded in a single batched RPC call
since 156ce433 (checks: Implement infrastructure for bulk diff
checks, 2021-07-29), we are now in a position to do the same for
commits: If we have an object quarantine directory, then we use the new
`ListAllCommits()` RPC with all alternate paths unset. This will thus
effectively only list commits of the object directory, which in a
quarantined environment points to the quarantine object directory.

Note that we still have to retain the old code to list commits via
a revision walk, given that non-push RPCs currently don't have an object
quarantine directory available. While this is slated to change soon by
creating manual quarantines, we must continue to support enumeration via
revwalks.

Changelog: performance