Action-Attentive Deep Reinforcement Learning for Autonomous Alignment of Beamlines