fix sim torque, prepare package for dreamer

[Armandpl]

The quanser sim doesn't expect a reduction ratio, fix that and try sim2real to gauge if the sim is anything close to reality. (Hoping it is so we can pre-train our dreamer world model)

• [x] take reduction ratio into account when computing the torque
• [x] investigate sim (integration?) divergence
  • take random actions and see what's up
  • it diverges often when training with dreamer, which is weird because it didn't do that at all when training with sb3?
  • if it does diverge, even with sb3, maybe add an integration step parameter?
  • ended up adding more integration steps, doesn't seem to be much slower tho i have't profiled it. lets run with it for now

Take the opportunity to also fix things we need for dreamer v3:

• 75

• [x] register env with gym
  • actually just use this syntax env = gym.make("furuta:FurutaSim-v0)
• [x] add dt obs not just yet, let's try without and improve from there
• [x] #60
• [x] fix the PID that is supposed to stop the motor
  • few weeks ago I broke the motor because the stop command was a step function from whatever speed to 0
  • we need to pass a slope and tune a pid to gently follow the curve. we should be able to stop the motor in 200-500ms gently
  • let's try it without it for now, motor seems to slow down fast enough
• [x] flatten history wrapper obs
• [x] check if the simulation and the pendulum rotate in the same way

[Armandpl]

alright so quanser sim diverges too, it seems that's due to the integration step being too big. I'm guessing it wasn't diverging before because of the speed limits i was enforcing. they were probably also preventing the sim from diverging.

• [x] let's add a smaller integration step and see how much longer it takes to train the agent in sim