Architectural and Inferential Inductive Biases For Exchangeable Sequence Modeling