Rethinking the Role of Scale for In-Context Learning: An Interpretability-based Case Study at 66 Billion Scale