On the Role of Initialization on the Implicit Bias in Deep Linear Networks