[http]::<tt was empty for bumped at 2/3 step configurations
(both fwd and tunnel) because HierarchyLogEntry::startPeerClock()
was not called for pinned connections.

Also adjusted [http]::<tt definition to account for the peer lookup
stage (which runs in parallel with connect attempts).

Also fixed [http]::<tt for peek-and-splice step 2/3 configurations when
a TLS error initiates switchToTunnel(). In this case, the logged
time covered only the interval up to the failed TLS attempt and
did not include the rest of the tunnel communication.

and a couple of polishing changes.

and a couple of polishing changes.

Also documented what should happen when the %code is used before
the timer stops. XXX: That is not what happens right now.

The only stopPeerClock() left is in FwdState::stopAndDestroy(),
which covers all cases (sucess, error) when communication
with the last peer terminates.

In this temporary approach the time is taken from HttpRequest::hier.
Also polished and simplified.

* do not dereference a possibly nil al->request
* optimized %<tt formatting code

We do not need this flag anymore since Stopwatch can track
multiple resumes and pauses.

It turned out that we cannot guarantee that condition when
switchToTunnel() was initiated by noteTakeServerConnectionControl(). In
this case, FwdState is destroyed after TunnelStateData, and therefore
the timer is still 'running' in TunnelStateData destructor.

thus addressing problems outlined in a693219.

The TODO looks irrelevant, since we do not (re)start peer clock
in TunnelStateData::startConnecting(), but still need to
reset hier.peer_reply_status there.

... documented in the previous branch commit XXX.

I am not sure this complexity is warranted, but storing a raw pointer to
a disappearing object (for some definition of "disappearing") did not
feel right either. This is the best solution I could find.

It is difficult to justify paying significant price for supporting
future hypothetical use cases, especially when we can add support for
other use cases (and make the class more complex) as needed instead.

They are no longer needed, but it is probably best to remove them when
refactoring this code to replace member assignments with member
initialization code. While at it, we should consider either adding
RefCount and CbcPointer methods to guarantee object existence (similar
to std::optional::value() and std::vector::at()) OR a global function
providing that same guarantees for any pointer-like object.

Context:
squid-cache#1828 (comment)

Context:
squid-cache#1828 (comment)

Context:
squid-cache#1828 (comment)