Investigating Mechanistic Interpretability in Large Language Models through Model Stitching Across Architectures