We discovered that on some weaker memory consistency architectures (read aarch64, etc) a read with acquire-fence is necessary in order to correctly observe an element being present in WorkQueue::push().

Unfortunately it is not feasible to add a regression test for this, as it very rarely manifests.

Diffing against https://github.com/openjdk/jdk25u/blob/master/src/java.base/share/classes/java/util/concurrent/ForkJoinPool.java

```
diff --git a/src/java.base/share/classes/java/util/concurrent/ForkJoinPool.java b/src/java.base/share/classes/java/util/concurrent/ForkJoinPool.java
index a7821921bc9..cdc5e0b9ea6 100644
--- a/src/java.base/share/classes/java/util/concurrent/ForkJoinPool.java
+++ b/src/java.base/share/classes/java/util/concurrent/ForkJoinPool.java
@@ -1297,7 +1297,7 @@ final void push(ForkJoinTask<?> task, ForkJoinPool pool, boolean internal) {
                     unlockPhase();
                 if (room < 0)
                     throw new RejectedExecutionException("Queue capacity exceeded");
- if ((room == 0 || a[m & (s - pk)] == null) &&
+ if ((room == 0 || U.getReferenceAcquire(a, slotOffset(m & (s - pk))) == null) &&
                     pool != null)
                     pool.signalWork(); // may have appeared empty
             }
```