The `gpg-agent` that could have been spawned here dies when it sees
it's socket disappear.

However, sometimes it seems like we fail to delete the homedir,
causing the `gpg-agent` to live on forever. We've noticed that the
deletion failed in
http://gitlab.com/gitlab-org/gitlab-foss/issues/36998: there was a
race condition during the deletion where `gpg-agent` would still be
modifying files while we've already called `FileUtils.remove_entry`.

This will attempt to delete the directory multiple times, at least 0.1
seconds apart. This is a naive way of trying to make sure
we clean up the homedir and count on `gpg-agent` to see that and make
itself go away.

On a web node we'll attempt for at most 0.5 seconds to clean up the
directory before failing. In a sidekiq process we'll attempt the
deletion for up to 2 seconds.

When the cleanup fails, we will now track that exception in
Sentry to gain some visibility.

This also adds counters for the creation and deletion of tmp
keychains, which we should be able to correlate to the nubmer of
zombie `gpg-agent` processes.