Uncertainty-aware Self-training for Few-shot Text Classification