Use more fair RL evaluation metrics: IQM, Optimality Gap, Median, Mean

[zuoxingdong]

Reference paper: Deep Reinforcement Learning at the Edge of the Statistical Precipice.
Github repo: https://github.com/google-research/rliable