Pareto-optimal clustering with the primal deterministic information bottleneck