Reaching quantum accuracy at the mesoscale in spatially disordered metal-organic frameworks by developing transferable interaction potentials based on active machine learning