Measure-valued Proximal Recursions for Learning and Control