Closes #3410

If we lose a header from http_SetHeader() we don't formally run out of
workspace prior to calling http_fail().

Refs #3410

And slightly move things around. There are interesting details worth
mentioning.

Better diff with the --word-diff --ignore-all-space options.

We actually changed behavior also for vcc_acl_pedantic=false

And a quick note on the workspace API plus homework for later.

Yes, I missed at least one varnishd change in my first pass.

Leaving out any change that is solely a bug fix on purpose.

The only user-visible change for this release.

I thought the slight change of behavior was worth mentioning, and that in
the varnishadm/CLI section it would stand out.

Whether the header was set by the backend or directly in VCL, it is now
possible to signal that a backend connection should not be added back to
the pool after a successful fetch with a Connection:close header.

Pooling such a connection would be counter-productive if closing the
session was requested by the backend itself, because it would then be
likely that reusing the connection would result in busting the extra
chance.

Setting the Connection:close directly in VCL can help mitigating against
a misbehaving backend.

Refs #3400

And amend the 7.0.0 mis-prediction.

Which isn't that many when you omit VCL.

In addition, soften the WS_Front() deprecation and keep the new WS_Id()
function internal.

Varnish uses a global per-probe timeout for backend probes. When reading
the backend response, Varnish tries to poll and read in a loop, until a
poll timeouts, the streams EOFs or there is an error.

The poll is supposed to timeout when the per-probe timeout ends. This is
currently setup so that `t_end` is the deadline for the probe, set when
the function starts, then the poll waits until `t_end`.

Previously, the poll timeout was never updated, and was always set to
`t_end - t_now` without updating `t_now`, which means that it was
effectively a between-bytes timeout instead of a proper per-probe
timeout.

This fixes this issue by updating `t_now` before updating the
`t_end - t_now` timeout so that the timeout passed to poll effectively
corresponds to a deadline of `t_end`.

See the issue fixed by this commit for more details.

Fixes: #3402

Give WS_Reservation() the same "always return non-NULL or assert"
semantics WS_Front() had because literally every single caller which
uses the ->f pointer fails to check for NULL.

Introduce a new internal WS_IsReserved() for use in the asserts
which check if a reservation is active, but which doesn't otherwise
care for the ->f pointer, and the single instance of non-WS code
which (possibly) legitimatly does variant processing depending on
the reservation state.

It has been a while since the built-in vcl_hit changed, and this
updates the user guide.

Since 6.0.1 we have had the corresponding change in TTL, grace, keep
change, and this is also mentioned.