Summary

Closes #1700. Drops full-suite runtime from ~85s to ~46s by
removing real-time waits from tests that don't need them on a
loopback socket. Behaviour under test is unchanged — only the
test setup is shortened.

Details

• quick_timing fixture applied to 13 register/announce-heavy
  tests (test_register_service_with_custom_ttl,
  test_logging_packets, test_ptr_optimization, test_qu_response,
  test_async_context_manager, test_integration in
  test_asyncio.py, test_asking_default_is_asking_qm_questions_after_the_first_qu,
  test_ttl_refresh_cancelled_rescue_query, both
  test_legacy_record_update_listeners,
  test_close_zeroconf_without_browser_before_start_up_queries,
  test_close_zeroconf_without_browser_after_start_up_queries).
  The fixture shrinks _CHECK_TIME / _REGISTER_TIME /
  _UNREGISTER_TIME from RFC 6762 §8.1/§8.3 production values
  (500ms / 225ms / 125ms) to 10ms each via patch.object, so
  nothing outside the with block is affected.

• _backdate_cache(zc, ms=1100) new helper in tests/__init__.py.
  Replaces real-time sleeps that exist only to satisfy RFC 6762 §10.2's
  "received more than one second ago" check (the cache flush logic in
  DNSCache.async_mark_unique_records_older_than_1s_to_expire):

  • test_async_updates_from_response: 3 × time.sleep(1.1) → 3 ×
    _backdate_cache(zc) (3.3s → 0.03s).
  • test_cache_flush_bit: await asyncio.sleep(1.1) → inline
    backdate of the TTL=1 records (1.1s → 0.0s).

  The helper iterates store.values() rather than the inner dict
  directly because DNSCache._async_add re-keys the dict with
  store[record] = record: when a new record with equal hash
  arrives, the key stays the original object but the value
  becomes the new record. Mutating the keys would update stale
  objects no one reads.

• Reaper-driven tests (test_reaper, test_reaper_aborts_when_done,
  test_service_browser_expire_callbacks) re-add records via
  cache._async_set_created_ttl(record, past, 1) so the cache's
  expire heap picks up the updated when = created + ttl*1000.
  Mutating record.created alone leaves the heap entry pointing
  at the original future expiration time and the reaper never wakes
  the record up. Combined with the _CACHE_CLEANUP_INTERVAL=0.01s
  patch these tests already had, the reaper now fires within ~10ms.

• test_service_info_async_request caps the post-unregister
  async_get_service_info to timeout=200 (the service is gone,
  the default 3000ms is pure overhead) and tightens the
  _is_complete=False TOCTOU-race loop from 3000ms to 1500ms.
  1500ms covers the initial _LISTENER_TIME + random_delay
  (≈220-320ms) and leaves plenty of margin for the loopback
  response to land before the loop times out.

• test_info_asking_default_is_asking_qm_questions_after_the_first_qu
  drops its forced-loop timeout from 1200ms to 500ms — enough for
  the QU query at t=0 and the QM query at t≈_LISTENER_TIME +
  max random delay (320ms).

• test_register_and_lookup_type_by_uppercase_name replaces
  time.sleep(1) with a 50×20ms poll on the cache (same pattern
  as test_sending_unicast in test_core.py).

• test_integration_with_listener_class drains the registrar's
  out_queue / out_delay_queue instead of time.sleep(0.5) for
  "multicast timers to expire" — same pattern as test: drop pending multicast responses before TOCTOU assertion #1701.

The remaining ~2-3s tests in the slowest-20 are intrinsic-timing:
_DUPLICATE_QUESTION_INTERVAL-bound queries
(test_we_try_four_times_with_random_delay,
test_get_info_suppressed_by_question_history,
test_get_info_partial), the multicast aggregation windows
(test_response_aggregation_timings, …_multiple), and the 1s
asyncio.run_coroutine_threadsafe(...).result(1) guard in
test_shutdown_loop. _DUPLICATE_QUESTION_INTERVAL is a cdef
constant in _services/info.pxd and _history.pxd, so it can't
be patch.object-ed in Cython mode.

Test plan

• Full suite passes: 335 passed, 3 skipped.
• All changed tests stable across five back-to-back runs.
• Pre-commit clean on the modified files (ruff, mypy, flake8,
  codespell, pyupgrade).
• Before/after slowest-20 comparison shows the targeted tests
  now drop off the list (most go from ~1.9-3.3s down to
  0.03-0.4s).