This needs a separate job. Also fix previous commit.

@bsdphk to be discussed / replaced

entire process group on timeout.

This still leaves the case where the process group leader exits
without a timeout but some of its subprocesses failed to stop.

Relevant to: #2469

unused since 992d5250

Please do not assume this to be complete since 5.2 release

Please do not even assume my XXXes to be completely referencing the missing bits.

Waving @slimhazard / Geoff from the ICE train and Berlin U-Bahn/S-Bahn

I wonder if maybe varnishd start should not return before accepting?

Ref: 6cb21b47
     #2382

Afterthought on #2405:

The usage pattern we want to get to is (pseudo vcl, don't take
the syntax for granted)

	sub vcl_whatever_client_sub {
		std.rollback(req);
		set req.<whatever>;
		...
		return (restart);
	}

The hash and storage controls should be no exception

Rewrite the CLI command acceptance to not use the VLU functions,
but a small state machine.

Unauthenticated CLI sessions have a hard limit of 80 char cmds
and cannot use the "<< nonce" construct.

... not for any vsm change.

Fixes #2430

Merges #2447

Fixes #2405 for restarts
Merges #2447

remaining TODOs (but not the scope of this ticket):

* Keep bereq attributes for bereq
* improved rollback

I see absolutely no reason why we shouldn't allow read access outside
vcl_recv

Merges #2447

This should have gone with d9a9ecd9

Somehow I was unable to see that the code did not actually do what I
thought (and documented) it would do.

Fixes #2470

a vav[0] is the error or NULL if ok. vsm_cmp_av considers NULL the end
of the av, so we would always comare vg and the av equal and begin insertion
of new vgs. This would lead to all vsm segments be removed and newly added.

Fixes #2470