Distributed Training

[LovelyBuggies]

Deploy each agent on a GPU, reward should be on CPU, as a follow up issue on #3

[LovelyBuggies]

First need a plan to see how to do it, better if we can keep the one-GPU training (like see whether the device number is smaller than the agent number, then automatically use this feature).

[LovelyBuggies]

I assume most of the task can be solved by less than 4 agents.

[LovelyBuggies]

Closed by #62