Gradient Descent for Deep Matrix Factorization: Dynamics and Implicit Bias towards Low Rank