LaMPP: Language Models as Probabilistic Priors for Perception and Action