Adversarial Robustness of In-Context Learning in Transformers for Linear Regression