A Unified Model for Zero-shot Music Source Separation, Transcription and Synthesis