A non-asymptotic distributional theory of approximate message passing for sparse and robust regression