Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation