- more carefully counting the running and exited workers

	- "< max threads running" is a warning, not alert

	- restored hash entry check against fd, but not if stale (-1)
	- removed membar after client bit not set (didn't help)
	- intial data send fail (before reconnect) is a warning (not alert)

	set -- false fd's tend to come after this, maybe data are
	stale
	- dump the data table before abort

	so that the old child process shuts down without delay

	- adopting varnishd.c's hack of calling nm -an to get symbolic
          names for stack traces

	- no connection pooling, always one connection per worker

	- monitor stats emitted one last time before abort

	because RPM is uncommonly stupid

	(trying to hint the compiler to use constant bit mask in hash)

	- reworked config params:
		- maxopen.scale can be an power of 2
		- maxdone, maxdata and qlen.goal not powers of 2