On the Effect of Initialization: The Scaling Path of 2-Layer Neural Networks