Change end-of-episode in CarRacing to termination as opposed to truncation

[RedTachyon]

There's been some heated debate about this, I'm still open to have my mind changed, but imo with the new termination-truncation split, this is a clear example of a termination as opposed to truncation.

Truncation is meant to represent hitting the "externally imposed" time limit as opposed to actually reaching a terminal state. When the car does the laps that it needs, the episode should be terminated.

To give some more information, I also added an info entry describing the reason for termination (whether or not the lap was finished)

This also requires a version bump, since implementations that correctly handle termination/truncation may converge to different policies

[pseudo-rnd-thoughts]

There are a couple of options which we need to plot the training curves for

1. Do nothing, keep the buggy truncation case
2. Fix the truncation testing to check that the agent truncates correctly after 2 (or more) laps (add the truncation lap number as an environment parameter, for later) - https://pypi.org/project/ufal.pybox2d/
3. Change to termination after a single lap

[Kallinteris-Andreas]

For reference, this is the initial discussion: #106

Here is my analysis, giving a truncation signal after finishing the LAP would imply that the environment is not over, that there would be more LAPs left to complete, while giving a termination signal would indicate that nothing matters after completing the LAP, for example crushing the car afterwords would be totally fine (realistically that is not going to happen).

Learning performance wise, I suspect both options would be really close regardless.

[pseudo-rnd-thoughts]

@RedTachyon Any chance you can run some training plots for this to measure the change?
Can you update the changelog for the environment, otherwise good with me

[RedTachyon]

Re: changelong - do you mean in the docstring? That's already done. The docs site page (I think) gets automatically generated from that.

Re: experiments - probably too much effort for the benefit. The impact would likely be visible in the qualitative behavior of the agent towards the very end of the race, and that's contingent on the implementation correctly accounting for the termination/truncation split