On the Use and Misuse of Absorbing States in Multi-agent Reinforcement Learning