Intelligent reflecting surfaces (IRSs) mounted on maneuverable aerial platforms to form aerial IRS (AIRS) relays represent a novel paradigm for large-scale downlink transmission in smart cities. However, the challenge of multivariate dynamic coupling hinders most existing studies due to high computational complexity and limited scalability. To address these issues, this paper proposes a soft-attention-actor-critic (SAAC) optimization framework that efficiently decomposes the joint optimization of multi-AIRS deployment, passive beamforming, and active beamforming at the base station into two sequential subproblems. The objective is to maximize average downlink spectral efficiency and service fairness, while minimizing deployment energy consumption. In the first stage, a conservative lower bound of spectral efficiency is formulated to guide multiple AIRSs toward near-optimal deployment positions. In the second stage, refined optimization is performed for both passive and active beamforming matrices. Furthermore, multi-head attention modules are incorporated into the critic and actor networks in each phase, enabling AIRS to adaptively attend to the observations and actions of other agents, and enhancing the ability to handle high-dimensional observation-action spaces. Extensive simulation results validate that the proposed SAAC framework consistently outperforms mainstream deep reinforcement learning baselines across diverse network conditions, highlighting its superior performance and scalability.