RFE: IdleTerminationTimeout directive to prefork/worker/event MPMs.

[notroj]

Builds on #529 by Alex/AT, which adds an IdleTerminationTimeout directive to
the prefork, worker and event MPMs: the server terminates itself once idle, so
that a socket-activated instance hands its listening socket back to the service
manager rather than sitting there.

The first commit is that patch unchanged. The rest are a test suite and fixes
for what it turned up:

• A server serving one request a second was counted idle at every sample and
  terminated. Idleness is now "nothing has happened since the last check".
• Connections held open by the event MPM occupy no worker, and were dropped.
  The open connection count is now published where it changes rather than only
  when a timeout queue expires, which also fixes ConnsTotal reading zero in
  mod_status with connections open.
• The countdown ran in maintenance cycles rather than seconds, so it expired
  early when children exited, and num_buckets times too fast with more than
  one listener bucket. It is now timed from a timestamp.
• The value was parsed with atoi(), so anything unparseable meant zero, that
  is, terminate as soon as idle.
• The timeout was not reset in pre_config, so it survived a reload which no
  longer configured it.
• The new shutdown message reused the SIGTERM message's APLOGNO.

test/modules/core/test_007_idle_termination.py covers all of the above, and
skips on MPMs which do not implement the directive. CI gains a prefork run of
the pytest suites, since nothing set MPM before.

Not done: documentation in docs/manual/. The shutdown is also still
ungraceful, which is safe now that it only fires with no connections open, but
that should be said in the docs.

Only the event MPM was exercised locally; prefork is covered by the new CI job.

🤖 Generated with Claude Code