Online stability of backpropagation-decorrelation recurrent learning