When Fairness Metrics Disagree: Evaluating the Reliability of Demographic Fairness Assessment in Machine Learning