Forward a consumed suppressed notification when a spinner has already stolen a task and other work remains queued. Otherwise the spinner can enter a blocking task while another newly scheduled task remains stranded in a queue with all other workers parked.

### Deadlock Scenario
When allow_spinning = true, idle worker threads spin for up to kSpinCount (5,000) iterations in MaybeGetTask() calling GlobalSteal():
1. While a worker is spinning (num_spinning > 0), any concurrent Schedule() or ScheduleWithHint() calls elide ec_.Notify(false) to minimize wake-up overhead, incrementing num_no_notification instead. The thread pool assumes the spinning worker will pick up the newly submitted task.
2. When the spinning worker finishes its spin loop, it calls StopSpinning(), which returns has_no_notify_task = true if any notification was elided.
3. If the spinning worker stole a different task t->f during its spin loop, unpatched code skipped the compensating steal and never notified a parked worker.
4. If all other worker threads in the pool are parked, the elided task sits stranded in its run queue. If the stolen task currently executing blocks waiting for the stranded task (e.g. at an XLA collective rendezvous / barrier), the thread pool permanently deadlocks.

### Fix
In MaybeGetTask(), when has_no_notify_task is true:
```cpp
if (has_no_notify_task) {
  if (t->f) {
    if (NonEmptyQueueIndex() != -1) ec_.Notify(false);
  } else {
    *t = GlobalSteal();
  }
}
```
Also acquire the queue publication in StopSpinning() and release in IsNotifyParkedThreadRequired() to pair queue publication with suppressed notifications.

### Testing
Added DelayingThreadEnvironment in threads_non_blocking_thread_pool.cpp to widen the interval between taking a task and retiring the worker's spinner, reproducing the lost notification under test_parallelism.

Backport of https://gitlab.com/libeigen/eigen/-/merge_requests/3038.

PiperOrigin-RevId: 987583048
Change-Id: I3336ed6ee95914579a54718e72f178865a3f27b5
2 files changed