[Pylot] Extend ERDOS agent driving policy

[Ionel Gog]

The ERDOS agent driving policy only follows fine-grained waypoints, and stops for pedestrians, vehicles and traffic lights (see https://github.com/erdos-project/erdos/blob/master/examples/pylot/erdos_agent_operator.py).

In order to make our agent more realistic we have to:

1. Remove dependency on depth camera to transform obstacles from 2D camera view into 3D world view. We could instead use two RGB cameras.
2. Explore more complex driving behaviors (e.g., lane change to overtake stationary vehicle). See https://github.com/ApolloAuto/apollo/tree/r3.5.0/modules/planning/traffic_rules for several examples of behaviors.
3. Ensure we can handle the situations in the Carla training data.

[Ionel Gog]

We've moved Pylot into its own repository.