Stabilised the models by forcing the production of k and epsilon to be positive

even when the dot-product of the Reynolds stress and the gradient of velocity is
negative.  Currently this is done by taking the magnitude, another alternative
would be to take only the positive part.