Implement new PyBullet backend for benchmarking purpose

[diegoferigo]

The current architecture can be recapped as follows:

• Robot: abstracts getting and setting robot values and references
• Task: works on a generic Robot object and implements the environment logic (how to process the action, how to form the observation, computes the reward)
• Task Wrapper (e.g. GazeboEnv): wraps the Task and steps the physics with the simulator

Pybullet is probably the main open-source alternative to the Mujoco simulator. There's a quite big community for robotic applications that chose pybullet. Waiting that Ignition Gazebo supports bullet as a backend (coming! :tada:), it might be a good idea adding the support of pybullet for benchmarking purpose. Only two components need to be developed:

1. PyBulletRobot: A new implementation of the Robot interface to interface with the bullet simulator
2. PyBulletEnv: A Task Wrapper that steps the physics using pybullet

If everything is developed correctly, the Tasks object should work on both simulators.

[diegoferigo]

Implemented in https://github.com/robotology/gym-ignition/pull/40.