Better diff with the --word-diff option.

We have two fallbacks in the absence of a .host_header field in the
backend, one of which is systematic. We can build a non .request probe
request with a single printf statement.

Please note that when the .host field is in the form "$addr $port" we
use that as the probe's host header. This behavior is not introduced by
this change but ultimately we might want to do something about it and
for example capture the first token from hosthdr with a txt.

In case of failure, I can see which client failed more easily.

rest of our test-programs are structured.

A lot of this is s/REQ_BODY_/BS_/g, followed by a reduction
pass to go from testing on specific status values to attributes
of the status value.

For instance the req.body caching code does not need to know if
how the transport sees this, chunked vs. eof for instance, it just
needs to know that there is a body and that we do not know the
length of it yet.

The transports own the "vocabulary", because internally H1 needs
to know the difference between BS_CHUNKED and BS_EOF in order to
setup the VFPs.

This worked only because REQ_BODY_INIT = 0

I hate C enums.

This reverts commit 54e7d45b.

However instead of reverting to the previous domain, I picked a
different one that I hope will not turn out controversial.

This is a result of a commit review between 6.2.0 and 6.3.0 and while
it should in theory be comprehensive, it is not. The upgrade notes in
particular are developer-centric, which may be a sign that none of the
changes require attention during an upgrade it could as well be an
oversight.

There is one item that I identified that should be mentioned in the
upgrade guide. I invite @nigoroll to give a shot at the VCL auto state.

In the absence of a fair warning, I extended WS_Reserve's grace period.

I omitted the -rcvbuf action on purpose, for the same reason why it is
not documented in vtc(7) on purpose: it's currently used in conjunction
with vmod_debug and we don't ship it.

While this is a milestone, I don't consider this work to be done so I
didn't remove the notes saying so.

We should also probably partially pick 53b4d2f5 on the 6.3 branch.

The Keybindings section is the only place mentioning differnt log levels
and one might assume that `-1` or `-j` behave the same as the default
ncurses interface.

The latter however does not show all counters by default as the single
action flags do, hence `varnishstat -f '*overload*'` will not show
`MAIN.sc_overload` by default, with `-1` however it does.

Fix grammar while here.

Restructured so that:

* 'Upgrading' is limited to work that has to be done to upgrade from
  a current deployment to the new version.

* 'Changes' is a comprehensive, user-level description of changes and
  new features.

Conflicts:
	doc/sphinx/whats-new/index.rst

They are however far from complete, and that's an UNDERSTATEMENT.

C-spec to understand what the SunOS compiler was whining about.

Strictly speaking it has a point, code is different from data is
different from const data, so mixing pointers the way we did was
not kosher.

We probably should have made the identifying argument a uintptr_t
instead of const void* to begin with to deal with this.

This reverts commit 2034138e.

We already know that SunCC will complain about those. Alternatively we
could use static variables addresses as cookies, but do we need extra
symbols?

Refs #3184

You never know what the next test case will look like...