Tracr: Compiled Transformers as a Laboratory for Interpretability