Towards learning universal, regional, and local hydrological behaviors via machine learning applied to large-sample datasets