Dynamic structured pruning and merging for real-time acceleration of reinforcement learning