With a tiny bit of polish here and there.

I'm purposefully leaving the shard direcor and vmod_std alone for now.

The previous ones were antiquated.

And bionic inherits the "oldest Ubuntu" 32bit setting.

The variable is independent of req.esi / resp.do_esi:

Whether or not the backend side prepares an object for ESI has nothing
to do with whether or not the client side processes it.

To avoid the maintenance of a downstream patch.

Like the one used by default on Debian Buster. Also, we don't want to
`set -e` by default when we use `false` on purpose...

There's technically nothing breaking there that require upgrade notes,
but it doesn't hurt, since it contains items related to both VNUM and
VRE breaking changes.

We could have shortlived_storage and req_body_storage parameters to stop
requiring Transient for those special cases.

Refs #3661

And simply require string parameters to define their tweaks.

Refs #3661

I believe I have laid out all upgrade-worthy sections, and started
filling them out.

Using groups[0].e turns out to be unreliable to print the suffix of the
subject string for a regsub operation. On Debian buster, with the help
of ASAN we can observe uninitialized memory through the remains of ASAN's
0xbe pattern that leads later to a complaint about an invalid pointer:

    runtime error: pointer index expression with base 0x6310000a0816
    overflowed to 0xbebf21cebec8c6d4

With a simple subtraction we can confirm the offset added to the base
address:

    0xbebf21cebec8c6d4 - 0x6310000a0816 = 0xbebebebebebebebe

To work around the possibility of an uninitialized ovector depending on
the pcre2 version, we initialize all offsets to PCRE2_UNSET and when we
encounter that value we capture a safe empty token.

This means that at the end of VRE_sub() we can no longer count on the
capture of the 0th group and revert back to using the offset.

Spotted by Martin.

I didn't bother adding references to other parts of the documentation.

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

Reported by Simon as a followup to #3683.

This makes use of the fact that Varnish will always send a session window
update immediately when receiving a data frame. The four rxwinup before
barrier sync on stream 0 matches up with the four txdata frames sent early
on the stream, and makes sure that the session thread has exhausted its
send window completely before opening up and starting to consume rxbuf
data by unblocking the client thread stuck in vcl_recv. From that point on
window updates will also be sent on the stream.

This reverts commit f6ee60cb.

A better attempt is on its way.

Fixes #3682

And while at it add coverage for `content-range: bytes */*` for good
measure.

Fixes #3683

Suggestions from Martin.