Skip redundant atomics on response completion - #2372

Every completed HTTP/1.1 exchange paid for atomic operations that had
nothing left to do, and every channel read paid for a fenced store:

- finishUpdate cancels the timeouts and done() cancels them again, so
  the second getAndSet always swapped null for null. Read the field
  first, as takePartitionKeyLock already does; the getAndSet still
  decides who cancels a holder that is there.
- done() offers the request back for release, but on the HTTP/1.1
  path it was handed to the encoder, so the CAS always failed. No
  state returns to OWNED_BY_AHC, so a plain volatile read can rule
  the CAS out; the CAS still arbitrates the race it was added for.
- touch() runs on every channelRead and stored the clock even when
  the millisecond had not moved. Skip the store when the value is
  already there; the field stays volatile and a skipped store would
  have written the same value.
- loadContent publishes "no more retries" with a release store. The
  only other access is an atomic increment, which reads the latest
  value regardless of the store's ordering, and the completion that
  follows happens-after it.

ResponseFutureCompletionBenchmark, JMH 1.37, JDK 17.0.2 on Windows,
3-5 forks: the second timeout cancellation 4.5-5.0 -> 0.45 ns, the
release of a handed request 5.1 -> 0.45 ns, touch 7.2 -> 3.9 ns,
cancel + done on a fresh future 225 -> 206 ns. Allocation unchanged.

Claude Code on behalf of Pavel Ptashyts

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Review feedback on AsyncHttpClient#2372. The release store for currentRetry saved
about 5 ns per response and would have been the only relaxed store in
NettyResponseFuture, so it goes back to a plain volatile set.

The read before the CAS in NettyRequest.release() leaves the CAS as
the only thing that keeps release() and markHandedToChannel() from
both claiming the body, yet no test raced the two: replacing the CAS
with a check and a store kept NettyRequestTest green. The new test
starts both calls on two threads in lockstep for 20,000 rounds and
checks that each body ends with exactly one owner. With the CAS
swapped for a check and a store it fails in the first rounds.

Local clean verify ran on JDK 17.0.2; CI covers JDK 11.

Claude Code on behalf of Pavel Ptashyts

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>