The interplay between implicit bias and benign overfitting in two-layer linear networks