Dual-Agent Reinforcement Learning for Adaptive and Cost-Aware Visual-Inertial Odometry