Learning Representations for New Sound Classes With Continual Self-Supervised Learning