LRP-based Policy Pruning and Distillation of Reinforcement Learning Agents for Embedded Systems