Rectifying Privacy and Efficacy Measurements in Machine Unlearning: A New Inference Attack Perspective