Observations passed to an agent must not be overwritten

[toslunar]

Some agents e.g. chainerrl.agents.DQN store a state argument of act_and_train without copying. This implementation doesn't always work well, e.g. if env.step constantly returns an array like env._state that has a fixed buffer.