non-fatal.

Fixes: #1414

Fixes: #1466

Fixes #1580

Fixes #1583

Fixes a regression from 33e6fe71

Fixes #1585

Also hide a might-be change until it's really implemented.

Thanks to dridi.boukelmoune@zenika.com for reporting.

This should bring us closer to the exit codes now documented in
varnishd.rst

Fixes #1572 for tools

This should bring us closer to the exit codes now documented in
varnishd.rst

Fixes #1572 for varnishd

Thanks to fgs for reviewing

testcases at the same time.

Spotted by:	UPLEX

After discussion with phk: exit(4) is a real exception, we should abort() whenever
possible.

The error codes patch is still to be committed.

Section 4.2 (Messages Headers) of RFC2616 defines field (header) name as case
insensitive, but the field (header) value/content may be case-sensitive.

http_GetHdrToken looks up a token in a header value and the rfc does not say
explicitly if tokens are to be compared with or without respect to case.

But all examples and specific statements regarding tokens follow the rule
that unquoted tokens are to be matched case-insensitively and quoted tokens
case-sensitively.

Fixes #1582

(This doesn't _quite_ close the race, but at least test c00034 is
much more stable now.)

tri-graphs.

Fixes #1566

Patch by:	fgs  (with minor modifications)

code could compile or not.

Fixes #1569

This mainly (only ?) changes things when we ask for an expression
of type STRING or STTRING_LIST.

Previously addition and subtraction would follow the general rule,
"first argument determines intial type" and that would make an
expression like this fail:

	set resp.http.server_port_foo = std.port(server.ip) + "_foo";

Because the addition tries to evaluate  "INT + STRING".

The idea was that people would rewrite this as:

	set resp.http.server_port_foo = "" + std.port(server.ip) + "_foo";

To make the first argument STRING and everything will then just work.

However, this is needlessly strict in the case where we are actively
desiring the expression to evaluate to STRING -- like above where
resp.http.* has type STRING.

With the new code, when an impossible addition is encountered, it
will look at the type requested of the expression, and if it is
STRING, convert the current subexpression to STRING and continue
adding.

A large number of subtests which are designed to fail started working
with this change, they have been fixed by not using *.http.* variables
as scaffolding..

You can still get into some weird corner-cases like the difference
between:

	set resp.http.foobar = req.ttl + 1s + "_" + req.ttl + 1s;
and
	set resp.http.foobar = req.ttl + 1s + "_" + req.ttl + 1;
and
	set resp.http.foobar = req.ttl + 1s + "_" + (req.ttl + 1s);

(Hint: The last one is the only one which does what you think)

Fixes #1579