Settling the Variance of Multi-Agent Policy Gradients