Updates Madras Env to support multiple single agent trainings, parallely

[buridiaditya]

What this PR adds to the Env

1. Updates Madras Env with the auto launch of torcs on a random vacant port.
2. Supports parallel agents.

Why this PR is required is

1. For parallel training of single agents.

Testing:

1. Clone the baselines and MADRaS repos.
2. Install editable copies of both using pip install -e ..
3. run python -m baselines.run --alg='ddpg' --env='Madras-v0' from baselines folder for single agent training.
4. run mpirun -np 4 python -m baselines.run --alg='ddpg' --env='Madras-v0' from baselines folder for parallel single agent training.

[rudrasohan]

The autostart function is also not working properly.

[buridiaditya]

• [BUG] pkill torcs kill all torcs instances.
• [BUG] update autostart.sh for auto navigation.
• [BUG] assign window names to torcs.

[rudrasohan]

@buridiaditya another one, when without vision training occurs after the env restarts the client waits for 100 sec to restart.

[rudrasohan]

There is one thing though as from a technical standpoint in both cases the without vision mode was working fine. But due to the frequent timeouts, I don't think the agent was able to learn anything. @buridiaditya is this a connection issue or that the simulator is literally insanely fast in the without viz mode.

[buridiaditya]

@rudrasohan Finished testing with MPI. Approve the PR if nothing else is left.