How does NeonDB avoid stale read issues when WAL logs are still in SafeKeeper and haven't been processed by PageServer?