Uncertainty Disentanglement with Non-stationary Heteroscedastic Gaussian Processes for Active Learning