Fixes #2946 in the sense that we want to treat a cacheable backend synth
like any other object and kill the stale object it replaces.

We do not replace the stale oc when
- abandoning the bereq
- leaving vcl_backend_error with return (deliver) and beresp.ttl == 0s
- there is a waitinglist on this object because in this case the default
  ttl would be 1s, so we might be looking at the same case as the
  previous

Thank you @mbgrydeland for the review and spotting a mistake.

Historically, all initializers called from VGC_Load() would handle
failures individually (return (1) from VGC_Load) or not fail at
all. Now that we got VRT_fail(), it makes sense to generalize.

The upcoming use case is to properly handle failures from
VRT_new_backend_clustered().

WS_ReserveSize() does not leave the workspace reserved when the
reservation fails, so WS_Release() must be called for retval > 0 only.

Besides the debug string, it is identical to WS_Reserve() except for

	assert(bytes > 0);

Follow-up #2967

This works with gcc 6.3.0 and clang 3.8.1-24

The test for __GNUC__ is deliberately simple and might not catch all
compilers which would potentially support deprecation marks. While more
specifics could be added, the aim is to raise awareness with developers
and we consider it quite unlikely that anyone does not compile with one
of the main stream compilers at all.

Seen fail on travis:

**   c1    2.7 === expect resp.status == 200
**** c1    2.7 EXPECT resp.status (200) == "200" match
**   c1    2.7 === expect resp.http.Age == 1
---- c1    2.7 EXPECT resp.http.Age (2) == "1" failed

To keep track of the number of times pipe_sess_max kicked in. This one
can be lazily updated.

Optionally, it can limit the number of concurrent pipe transactions as
they each monopolize one worker thread.

This allows users to keep track of the number of ongoing pipe
transactions at any time.

Refs #2948

Closes #2948

Fixes	#2959

With 32ef5dce, this test also cancels
backend requests and thus may prevent the server from finishing writing
the reponse body

Ref #2964

I had overlooked the fact that for a canceled request we might only read
parts of the body.

Interestingly, this is only exposed on the vtest ARMs.

Ref #2964

This is the intent that was not obvious to me when I first reviewed this
change.

Ref #2963

as requested by @Dridi

for anything but return (deliver)

Fixes #2963

... to un-confuse the interface

Notes on changes from WS_Reserve():

* Removed the first WS_Assert because all we change is ws->r and we got
  a specific assert on it.

* it follows from PAOK(ws->e) && PAOK(ws->f) in WS_Assert() that
  PAOK(ws->r) && PAOK(b), so we remove the PRNDDN()

Closes #2967

Yet having the changelog in one place is useful

    Missing HEAD method usage flag reset between two requests.

    If an HTTP client request with a method different from HEAD followed
    another one with HEAD method, the former was still flagged as being
    using the HEAD method because of a missing HEAD method usage flag
    reset before building an HTTP client request.

Author: Frédéric Lécaille <flecaille@haproxy.com>

    Reinitialize cyclic barrier only once per cycle

    It was done for all threads syncing on a barrier. The one reaching the limit but
    also all waiting threads. So it could be reset by a thread while others have
    already started another round, leading to an deadlock (because some increments
    will be lost then).

    Now, cyclic barrier is reset for a new round by the thread reaching the limit
    and only this one.

Author: Christopher Faulet <cfaulet@haproxy.com>

See #2967 for discussion

... where we would need error handling

to identify the vmod by its address in other places like vcc-assigned
VRT_priv_* ids