Stagewise and interpretable emergence of inside-outside computation in Transformer-based language models