On the Interplay Between Sparsity and Training in Deep Reinforcement Learning