Fix race in shared-stream overflow test that could hang the suite

[potiuk]

The scheduled canary flaked on
test_ack_mode_queue_full_during_fanout_does_not_break_iteration with a
TimeoutError (e.g. the LowestDeps ARM run).

Root cause is a race in the test, not the production code. The test
pre-fills trigger 1's subscriber queue with a sentinel so the background poll
task hits QueueFull and overflow-fails it at fan-out. But it then started
draining s_full immediately, so depending on event-loop scheduling the drain
could consume the sentinel first — leaving the queue no longer full when the
burst was fanned out. The put_nowait then succeeded, the overflow under test
never happened, and the iteration blocked forever (surfacing as the
TimeoutError). It passed most of the time only because the poll task usually
won the race.

Fix: synchronise on the overflow — wait until trigger 1 is recorded in
_failed_subscribers before draining s_full. The test is now deterministic
(passes 6/6 locally in ~1s instead of hanging), and the production overflow
path is unchanged.

---

Was generative AI tooling used to co-author this PR?

• Yes — Claude Code (Opus 4.8)

Generated-by: Claude Code (Opus 4.8) following the guidelines

[potiuk]

Almost identical #68651 merged before.