[epompeii]

Stacked on #1490.

While a maintenance unit waits for the runner's job lock, its main process is flock. When an unattended upgrade replaces util-linux or libc6, Ubuntu 24.04's needrestart flags the waiting unit and runs systemctl restart on it. That restart waits for the new start, the new start waits for the lock, and the upgrade run that holds the lock is an ancestor of needrestart, so nothing ends the wait.

cargo ops host now writes /etc/needrestart/conf.d/zz-bencher-quiet.conf, made from the list of maintenance units, which matches each unit by its exact name and has needrestart defer its restart. A deferred restart costs nothing for these oneshot units, since the next run starts the new binaries. cargo ops audit checks the file like every other file host writes.

• The file covers every unit under the job lock, the 2 apt units too. needrestart already defers those by default, and an apt-daily.service that waits for the lock has the same problem, so the file follows the one list.
• host writes the file on a host without needrestart as well, where it does nothing, so a later install of needrestart is covered. needrestart reads its configuration on each run, so nothing reloads it.
• A new test loads the generated file through perl the way needrestart does and checks that each unit is deferred by its exact name alone, that its timer and similar names are not, and that needrestart's own defaults stay.
• Checked on Ubuntu 24.04 with needrestart: without the file, the restart of a waiting logrotate.service hangs; with the file, needrestart defers it, exits 0, and logrotate finishes once the lock is free.