prometheus

Previously prometheus/alertmanager were unconditionally reloaded every 10
minutes. Reloading alertmanager when it is about to deliver an alert
notification seems to make it account successfull notification deliveries as
failed, increasing alertmanager_notifications_failed_total. So reloading every
10 minutes caused spurious AlertmanagerNotifications alerts in our setup.

Also config updates are now atomic regardless of whether /tmp is a tmpfs.