Unpacking SDXL Turbo: Interpreting Text-to-Image Models with Sparse Autoencoders