It used to be that we'd call http processing on a filedescriptor and
then that was that.  Then we added keywords to do new accepts in the
server and suddenly the calling code closes a wrong filedesc which
belongs to somebody else and things get really inconvenient fast.

This made all test-cases which use the "accept" server directive flakey.

test failures seems to be timeouts on stressed machines.

Teport timeouts more clearly.

two separate timestamps just happen to be the same, just make them
the same.

VSL record.  Still not quite sure how they happen, but we shouldn't hang
no matter what.

the same address again, even with hinting.

pretend nothing else happened.

this should solve issues where the c2-v1 connection times out before
it ever got used.

fd's early in the case of a timeout.

Fixes	#1023

Apache (and nginx) uses 8k, so use the same default size to avoid
people being surprised by more 413 responses than necessary.

Fixes: #1016

Thanks to cleberjsantos for initial patch which this is based on.

Fixes: #976

are no lingering bits.

Convert an assert to a failure.

Fixes	#1022

steps the clock...

Fixes	#1022

Make pipe mode use the acct_tmp like everybody else.

Make one single thread, which creates thread pools, and since that
is not particularly ardous work, also have it maintain the stats
gauge of number of sessions queued.

(Removing pools should now be possible, but the code is not there yet.)

Each pool has its own herder thread, which creates and destroys
threads as required, and updates stats for dropped and queued
sessions.

Sanitize VSC fields associated with this area.