Fix awaitEmpty() race in InFlightWindow

awaitEmpty() had a TOCTOU race between its error check and
in-flight count check. When the I/O thread called fail()
(setting lastError) and then acknowledgeUpTo() (draining the
window to zero) before the flush thread was scheduled, the
while loop exited on the count condition without re-entering
the body to call checkError(). This caused flaky failures in
testErrorPropagation_asyncMultipleBatchesInFlight.

Add a final checkError() after the while loop. Correctness
relies on the happens-before chain through the volatile
highestAcked field: the I/O thread's lastError.set() precedes
its highestAcked write, and the flush thread's highestAcked
read precedes its lastError.get(), so the error is guaranteed
to be visible.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: mtopolnik

core/src/main/java/io/questdb/client/cutlass/qwp/client/InFlightWindow.java

@@ -287,6 +287,11 @@ public void awaitEmpty() {
                 }
             }
 
+            // The I/O thread may have called fail() and then acknowledgeUpTo()
+            // before this thread was scheduled, draining the window while an
+            // error is pending. Check one final time after the window is empty.
+            checkError();
+
             LOG.debug("Window empty, all batches ACKed");
         } finally {
             waitingForEmpty = null;