In recurrent-style training, it stores data in the format of [batch_size, stack_num, ...]. And if we want to get the first hidden state, batch[:, 0] is the best way instead of manually and recursively get the value in the batch. (#19)

Also with the multi-agent training scenario (someone is working on it), the batch-data will store in the format of [env_num, agent_num, ...]. If we want to get data from a specific agent's perspective, the best way is batch[:, agent_id].

@duburcqa