For now I am satisfied that the delay is gone and will read up some more on tcp_nopush.
For the future: Is there any information on why the combination of tcp_nopush and proxy_cache_background_update create the delay and not the STALE response you get when the backend ist down and proxy_cache_background_update is off?