Cellular-connected unmanned aerial vehicle (UAV) swarm is a promising solution for diverse applications, including cargo delivery and traffic control. However, it is still challenging to communicate with and control the UAV swarm with high reliability, low latency, and high energy efficiency. In this paper, we propose a two-phase command and control (C&C) transmission scheme in a cellular-connected UAV swarm network, where the ground base station (GBS) broadcasts the common C&C message in Phase I. In Phase II, the UAVs that have successfully decoded the C&C message will relay the message to the rest of UAVs via device-to-device (D2D) communications in either broadcast or unicast mode, under latency and energy constraints. To maximize the number of UAVs that receive the message successfully within the latency and energy constraints, we formulate the problem as a Constrained Markov Decision Process to find the optimal policy. To address this problem, we propose a decentralized constrained graph attention multi-agent Deep-Q-network (DCGA-MADQN) algorithm based on Lagrangian primal-dual policy optimization, where a PID-controller algorithm is utilized to update the Lagrange Multiplier. Simulation results show that our algorithm could maximize the number of UAVs that successfully receive the common C&C under energy constraints.