Diffusion Lens: Interpreting Text Encoders in Text-to-Image Pipelines