Unveiling the Potential of Robustness in Selecting Conditional Average Treatment Effect Estimators