Suggested by Federico.

While working on libvarnishapi I couldn't get the VUTs to fail on
an INCOMPL() statement, this was caused by the recent changes to
the `shell` command.

The original test was:

    shell -err {
            varnishlog -n ${v1_name} -d -g raw |
            grep "Should not happen"
    }

But it wouldn't catch errors, like for example varnishlog failing to
connect to the shmlog.

Instead of crashing the varnishtest sub-process with an assert,
check that both `vcl.load` and `vcl.use` pass or fail when running
the varnish command for `-vcl` and `-errvcl`. `vcl.use` may fail
the warmup step.

See 9849fd06 for how it was solved.

We also don't want the test to pass if varnishlog failed.

This is an (intentional) regression from 6f9595e6,
but I don't see why we should prevent stringifications of DURATION.

Related to #2101 #2102

we can ditch the bogus return values and some decorations.

Right now this only affects failures in vcl_init{}

it more than once.

which means not in vcl_backend_error{}

Fixes: #1865

with their length.

Explains some of #2117 and friends

I'd really like to come back to https://lists.gt.net/varnish/dev/39240 somewhen.

We should not leave the workspace sizing calculations to the admin based on
incomplete information. The admin should only need to set the "space available
to VCL".

Varnish doesn't like empty hosts in backend definitions

It actually fails to build with recent enough GCC or Clang.

- 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