Revisiting Wright: Improving supervised classification of rat ultrasonic vocalisations using synthetic training data