Reducing Oversmoothing through Informed Weight Initialization in Graph Neural Networks