RLx2: Training a Sparse Deep Reinforcement Learning Model from Scratch