Eau De $Q$-Network: Adaptive Distillation of Neural Networks in Deep Reinforcement Learning