Bayesian Learning of Other Agents' Finite Controllers for Interactive POMDPs