PSLA: Improving Audio Tagging With Pretraining, Sampling, Labeling, and Aggregation