fix: resolve lost-wakeup races in InMemoryTaskStore.wait_for_update

[blackwell-systems]

Summary

Two races in InMemoryTaskStore.wait_for_update that can cause indefinite hangs when polling task status.

Race 1: Concurrent waiters

wait_for_update overwrites _update_events[task_id] with a fresh event on every call. If two clients poll the same task, the second overwrites the first's event. notify_update sets only the latest event, so the first waiter hangs forever.

Fix: Use a list of events per task_id. Each waiter appends its own event. notify_update sets and removes all events for the task.

Race 2: Notify before wait

If update_task calls notify_update before any wait_for_update is active, the signal is lost (no event exists to set). The next wait_for_update creates a fresh event and waits for an update that already happened.

Fix: Track pending updates in a set. notify_update adds to the set when no waiters exist. wait_for_update checks and consumes the pending flag before creating an event.

Reproducer

Both races are reproducible with a 30-line script (included in issue #2535). Output before fix:

Race 1: concurrent waiters
  waiter a: HUNG
  waiter b: woke
  FAIL: lost wakeup

Race 2: notify before wait
  FAIL: signal lost, waiter hung forever

After fix:

Race 1: concurrent waiters
  waiter a: woke
  waiter b: woke
  PASS

Race 2: notify before wait
  PASS

Test plan

• Reproducer script confirms both races are fixed
• All 217 existing task store tests pass (pytest tests/experimental/tasks/ -v)
• 0 regressions

Fixes #2535