CARL: Congestion-Aware Reinforcement Learning for Imitation-based Perturbations in Mixed Traffic Control