Move the clean up call to HSH_DeleteObjHead inside of the hash deref
methods that require it, instead of having the outside (single) caller do
it based off the return value. This just cleans up and makes the logic
more transparent.

This is caught by

	assert(lr == HSH_HIT || lr == HSH_GRACE);

in cnt_lookup()

Closes #2743

Conflicts:
	doc/changes.rst

doc/changes.rst was considered binary in master, it won't be in 6.0 so
there will be no need to pick a1d2db69
too.

Refs a1d2db69

in particular, move logging

they are the same case

Now that we got hit-for-pass (hfp) and hit-for-miss (hfm), we should
name the latter consistently to avoid unnecessary confusion

Conflicts:
	lib/libvarnishapi/vsl_arg.c

For some reason it was missing errno.h...

A vmod may return a null IP. This relaxes the check in VRT_acl_match to
fail the transaction instead of crashing.

Refs #2842

Inclused re-declaration fixes for some places which I got wrong in the
past.

Conflicts:
	bin/varnishd/cache/cache_objhead.h

This is a squashed commit including lint removal and formatting
suggested by @nigoroll.

The vmod name got lost in e7c04974

example before/after diff:

- new xshard_param = shard_param()
+ new xshard_param = directors.shard_param()

Partially addresses #2846. I've avoided mentioning NULL on purpose as
it's not really meaningful nor accessible from VCL. ymmv.

Pedantic back-port note: those are acronyms.

Add tests for vcl_keep and move vmod_so_keep together.
Reported and diagnosed by jw @ irc.

Enabled when using varnishtest -L.

this would have caught 7395ecae

Other polish.

$Module description, and suggest people to do it that way.

Make reporting syntax errors easier.

This adds rate limiting to varnishncsa and varnishlog.

Rate limiting is done on a per-transaction basis, respective to the
grouping mode selected. I.e. for -g request the limit will apply on a
per-request basis, -g raw on a per-record basis, etc.

Rate limit is specified as -R <limit>[/duration]. Default period if not
specified is seconds ('s').

This adds a ctx->handling test at the beginning of any compound
statement, to catch the cases where VRT_fail was invoked as part of an
if test condition.

Fixes: #2840

This may break out of tree code not respecting include order of vdef.h
first (or via cache/cache.h). It's trivial to fix and forces consumers
to follow the tracks.

Refs #2791

Seen once, and failed to reprocuce (yet), needs to be understood
better

from vmod_blobdigest tests/usage.vtc:

**** v1    1.8 CLI RX|No panic to clear
***  v1    1.8 debug|Info: manager stopping child
***  v1    1.8 debug|Debug: Stopping Child
**** v1    1.9 vsl|          0 CLI             - EOF on CLI connection, worker stops
***  v1    2.8 debug|Info: Child (174227) ended
***  v1    2.8 debug|Info: Child (174227) said Child dies
***  v1    2.8 debug|Debug: Child cleanup complete
***  v1    2.8 debug|Assert error in MCH_TrackHighFd(), mgt/mgt_child.c line 203:
***  v1    2.8 debug|  Condition(fd > 0) not true.
**** v1    2.9 STDOUT poll 0x10
**   v1    2.9 WAIT4 pid=174215 status=0x0086 (user 0.914949 sys 0.567445)
*    v1    2.9 Expected exit: 0x0 signal: 0 core: 0
---- v1    2.9 Bad exit status: 0x0086 exit 0x0 signal 6 core 128
*    top   2.9 failure during reset

On the 6.0 branch the regression appeared in ca76ef57 instead.

Side note

	#define CLOSE_FD_UP_TO	mgt_max_fd

exposes that we are currently missing to track some fds.

The name CLOSE_FD_UP_TO implies that the value is included.

Also this is the semantics of MCH_TrackHighFd() and relevant for
 #define CLOSE_FD_UP_TO mgt_max_fd

Ref 3c0b8768