Write an API for consistent training process

[Kytabyte]

It's a bit awkward to write a totally different script while training on another game or use another network.

The observation shape may not fit the network structure. We have to reshape them manually during writing a train.py.

We should write an API somewhere to check the observation shape and the input shape of a network, and a wrapper function to make them consistent wherever needed.

[Kytabyte]

We should write an Agent base class to wrap up RL algorithms and models.