When Active Learning Fails, Uncalibrated Out of Distribution Uncertainty Quantification Might Be the Problem