MADial-Bench: Towards Real-world Evaluation of Memory-Augmented Dialogue Generation