Evaluating Text-to-Image Diffusion Models for Texturing Synthetic Data