in "Upgrading to 5.1".

Also make an optimistic statement about the compatibility of VCL
written for 5.0 with 5.1.

Also install nghttp2 so a02022.vtc has a chance to run.

While at it, get rid of the hard-coded package name.

Refs #2251

for the gcov vtester to finish in 60 seconds.

So far this what I find in the commit log from varnish-5.0.0 until
the beginning of 2017. Some "biggies" still to be mentioned are:

* varnishd command-line options
* VCl type conversions, stringification etc.
* ranges?

It closes a race for barriers because the list could be updated by the
top thread while walked by another thread. It also saves useless lookups
by rejecting names earlier.

This is not the case for servers because a) they have proper locking of
the servers list and b) there is the dispatch special case. It would
otherwise be harmless to do it for the server too because dispatched
servers don't depend on user input.

We aren't building a vrt_test binary, it's a script. We also don't need
to distribute vrt.c, the cost of a copy doesn't matter and it ensures
that this file is always in the $(buildir). I also cleaned other wrong
assumptions regarding $(srcdir) vs $(builddir).

This reverts commit e706b22a. The real
problem was solved in d906981a. The
extension adds noise considering the make rules below don't comply with
the target name.

With a more assertive test this time.