[rllib] Exception for termination at an environment's time limit

[zigui-ps]

Describe your feature request

Time Limits in Reinforcement Learning (paper: https://arxiv.org/abs/1712.00378)

IMO, For time-unlimited tasks, an agent getting 0 rewards at the beyond of time horizon can decrease RL performance.
I suggest making an option to select a time limit exception in rllib.