Add road regeneration and pretraining

[araffin]

• upgraded stable-baselines
• add seeding and road generation
• add pretraining using behavior cloning
• add data-augmentation

TODO: tune pretraining

[vanmil]

Wouldn't it be nicer to use the gym register method instead of removing it?

In my own code I'm now using

from gym.envs.registration import register
register(
    id='donkey-v0',
    entry_point='donkey_gym.envs:DonkeyEnv',
    timestep_limit=2000,
)

and then in the code part where I learn the RL algorithm, I pass arguments to the gym.make command to init the env with the desired simulator level and vae object (or None if not using vae) like env = gym.make('donkey-v0', vae=vae, level=3)

By the way; great repo! ;)

[araffin]

Good point, in fact this is a recent addition to gym, this feature was missing for a long time, hence i was not using the make function.