On the Optimality of Binning for Distributed Hypothesis Testing