Questioning the Coverage-Length Metric in Conformal Prediction: When Shorter Intervals Are Not Better