Benchmarking Scalable Predictive Uncertainty in Text Classification