Are Uncertainty Quantification Capabilities of Evidential Deep Learning a Mirage?