Damped Anderson Mixing for Deep Reinforcement Learning: Acceleration, Convergence, and Stabilization