master: Fix service { idle_kill } to work better on busy servers
The previous behavior was to kill a process once it had idled for idle_kill
seconds. However, on a busy server the new connections are picked up somewhat
randomly by all the idling processes, so there's never any single process
idling for a long time. This effectively prevents the idle_kill from killing
any processes, even if there are unnecessarily many of them.
The new behavior here tracks the lowest number of idling processes during
idle_kill time interval. Then it kills that many processes. If the load
stays the same, this should shrink the number of processes to the number
that is needed to handle the load, but no more.