Fixes: #2258

We may see it after a write on a socket which has been closed by
the remote end.

Fixes #2267

triggers this bug for H2.  (H2 will be added to the testcase as part
of the fix.)

(Yes, obviously I managed to break 5.1 with my last commit before
the tag...)

Closes #2252

Refs #2251

For consistency.

Fixes #2251

because of a H1 assumption in transport-independent code.

the test timeout (-t) through a layer of makefiles (aka: make check).

This is not a supported platform, and we can still collect test logs for
failing jobs.

This reverts commit c2070784, we'll
stick to varnish modules. I also updated a commented-out occurrence
of "VCL modules" in params.h, we can probably get rid of these
comments.

The varnish-cli will however still report when a parameter is not
available. Spotted by Geoff.

For future reference, I browsed the changes for this development cycle
easily with:

    git log --patch --reverse varnish-5.0.0.. -- include/vrt.h

As suggested by Geoff, it probably deserves a short description.

This reverts commit 4c9f9e08. The
generated HTML and troff code wasn't valid anyway. That would ruffle
another OCD of mine.

This reverts commit 100adab2. In-place
sed on FreeBSD isn't compatible with GNU sed. It could be fixed with a
temp file, but meh.

According to vtest the hack fails on FreeBSD i386, so either it fixes it
or I revert the whole thing.

Apparently rst2man and sphinx don't have the same opinion on what an
option is or isn't. Looking at the RST spec it doesn't seem to allow a
question mark [1] in an option list. Among the various kind of options
we can see:

    Short POSIX options consist of one dash and an option letter.

To work around that I introduced a crude hack using sed. It satisfies
my OCD but I'm not even remotely proud of that.

[1] http://docutils.sourceforge.net/docs/ref/rst/restructuredtext.html#option-lists