Harnessing Synthetic Datasets: The Role of Shape Bias in Deep Neural Network Generalization