Behind the Myth of Exploration in Policy Gradients