#+title: AePPL's intermediate representation should be a first-class citizen

Many model transformations (including marginalization) happen at the measure level, AeMCMC needs the representation in terms of measure to build samplers. We should thus have a user-facing function that takes an Aesara graph with random variable and returns a graph with the corresponding measures.

joint_logprob can thus be a high-level function that delegates to this function and another one that computes the conditional density from that representation, and adds transformations like marginalization between these two steps.