Refactor parse_args()

[vwxyzjn]

This PR refactors a few arguments in parse_args():

1. This PR moves the gym_id, learning_rate, total_timesteps move down to algorithm-specific arguments. Currently when doing a file diff between c51_atari.py and ppo_continuous_action.py, we have the following screenshot:
   
   However, it's more desirable to all of the differences in one place, as shown below:
   

2. This PR also renames the gym-id to env-id. The reason is that as we adopt different environments, some of them don't necessarily have a gym-id but they have an environment name. For example, procgen has the env-id but not gym-id.

Curious about your thoughts @yooceii @dosssman on whether this is necessary.

[gitpod-io]

[dosssman]

I concur on the gym-id to env-id change, as the former is more general indeed.

For the position of the arguments in the argparser list, since some algorithms are kind of made for speicfic tasks (SAC -> Bulet / Mujoco), it does make sense to put it in the algorithms specific arguments. learning-rate too, because in SAC for example, there are actually 2 learning, so it would make the parameterization cleaner.

While total-timesteps is also general, I guess different algorithms training for different amount of steps could be possible (for example Atari's required 10M + vs Bullet's 1M) can be used as justification for this change.

In any case, it is all good on my side.

[vercel]

This pull request is being automatically deployed with Vercel (learn more).
To see the status of your deployment, click below or on the icon next to each commit.

🔍 Inspect: https://vercel.com/vwxyzjn/cleanrl/AAxRK8ytYUqJoht2izi8wtmoQzwA
✅ Preview: https://cleanrl-git-refactor-arg-vwxyzjn.vercel.app

[vwxyzjn]

The matplotlib utilities might break but that's ok since we are re-building Open RL Benchmark anyway #115.