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.

Spotted by Flexelint.

With this vav_test needs to link against libm, but it's simpler to tell
libvarnish to link since it's the one shipping VNUM in the first place.

Fixes #3645

With the hex table suggested by fgs.

That's just another case of falling back to regular delivery if the VCL
mistakenly attempts to set up a range filter.

That is, only when http_range_support is on, which is the default.

Refs #3246

While at it avoid reusing the same client name twice to help them stand
out in the test log.

Ref #3661

Spotted by Flexelint.

The larger the response we hit, the more likely we are to hit an ongoing
fetch and since we buffer transaction logs it can become really tedious
to infer whether the backend transaction may have already been overrun
or not flushed yet. At least, such a heuristic was impossible with a VSL
query until now.

It's a thin wrapper on top of vre_capture() that returns the number of
groups captured during a match.

Now that txt appears in vre.h we need vdef.h anywhere the former is
included. There was one generated file where that wasn't the case.

Closes #3655

And add a bunch of assertions in addition to the implicit ones from
pdiff().

There are other components than the cache process that could benefit
from it, in particular libvcc. The relationship with vas.h is somewhat
unfortunate but that also centralize the Tcheck() logic in pdiff() and
doesn't actually require vas.h unless the macros are used, in which
case it's almost guaranteed that the calling code already included
vas.h in the first place.

The benefits should outweigh the drawbacks.

My absolute bad on this copy-pasta mistake.

Only from a git clone, and only from the root directory, which should
leave alone distcheck and builds from release archives.