Inferring fracture aperture distribution at the EGS Collab Experiment 1 testbed through a deep learning accelerated Bayesian approach