- Try to distinguish between DURATION, REAL and INT automatically.
- Explicitly mention which operands are possible when adding.

Related to #2101.

This reverts commit a4e07fc5.

It still fails on FreeBSD and SunOS.

If we receive a cancellation during pthread_cond_(timed)wait() we will
own the mutex on the way out. Avoid locking it twice, otherwise we
will deadlock.

(Ab)use the bound address used for the bad_backend macro to expose a
random port and avoid the dreaded collisions.

The port may not be hardcoded, it is still available for other processes
to steal during the race between `varnish v1 -stop` and the next `varnishd`
execution.

Fixes: #2200

Because it uses hard-coded ports, it often appears as failing in vtest
so now it should succeed if the address wasn't available.

Spotted by Martin, when a VMOD relying on varnish.m4 is built with
--prefix in the configure command-line, a pseudo hard-coded vmoddir
is assumed:

    $libdir/varnish/vmods

Instead, it now asks pkg-config for the vmoddir relative to the
libdir. It worked in all cases with upstream Varnish Cache, but
downstream modifications of the vmoddir would then be ignored.

Fixes #2197

Change some tests to use the new parameter and more coverage.