This review synthesizes recent advances in hybrid control architectures for autonomous multi-drone swarms, with particular emphasis on approaches that integrate deep reinforcement learning (DRL) with model-based control and safety-aware mechanisms. The review examines swarm coordination as a constrained decision-making problem under nonlinear quadrotor dynamics and graph-based interaction topologies. It surveys major hybridization strategies, including DRL combined with classical controllers, planning-learning frameworks, swarm-intelligence-assisted optimization, and safety-enhancing techniques such as control barrier functions and runtime shielding. Drawing on literature published between 2018 and 2026, the paper evaluates the current state of the field and highlights key trends, strengths, and limitations of existing approaches. Particular attention is given to the challenges of achieving safe decision-making, scalable multi-agent coordination, and reliable deployment in real-world environments. The review identifies three major research gaps: the lack of scalable frameworks that combine learning with formal safety guarantees, limited solutions for robust simulation-to-reality transfer of collective behaviors, and insufficient consideration of hardware constraints during training and evaluation. To support more consistent assessment of hybrid swarm architectures, the review discusses evaluation perspectives centered on safety, reliability, efficiency, scalability, and deployment feasibility. Finally, it consolidates emerging best practices from the literature and outlines future directions, including verifiable hybrid control synthesis, scalable multi-agent reinforcement learning, hardware-aware policy design, human-swarm interaction, and standardized benchmarking. The review provides researchers and practitioners with a structured roadmap for developing safe, scalable, and certifiable autonomous drone swarms.