Reward prediction-errors weighted by cue salience produces addictive behaviors in simulations, with asymmetrical learning and steeper delay discounting