Beyond accuracy: quantifying trial-by-trial behaviour of CNNs and humans by measuring error consistency