A Posteriori Learning for Quasi‐Geostrophic Turbulence Parametrization