Self-Supervised Music Source Separation Using Vector-Quantized Source Category Estimates