Global Convergence Analysis of Deep Linear Networks with A One-neuron Layer