Text this: Rethinking spatial-temporal contrastive learning for Urban traffic flow forecasting: multi-level augmentation framework