The Double-Edged Sword of Implicit Bias: Generalization vs. Robustness in ReLU Networks