Towards a Mechanistic Interpretation of Multi-Step Reasoning Capabilities of Language Models