SMARLA: A Safety Monitoring Approach for Deep Reinforcement Learning Agents