Scalable Inference for Logistic-Normal Topic Models