Restructuring MADRaS into a Multi-agent MDP

[Santara]

Pending work: [Do not merge with master until the list is empty]

1. Add randomization features
2. Add training code for rllib
3. Add done handling for the case when the environment done is set when all the agents hit done.

[rudrasohan]

Updates:

For bringing the team up to speed with the latest developments and future actions.

Required:

• [ ] Simplify code structure. Merge v1 and v2 version of the files (For eg., reward handler, obs handler, and done handler).
• [ ] Add randomization and experiment automation features to v2.
• [ ] Make single agent a special case of Multi-Agent structure.

Additional (To be covered in either this or a separate PR):

• [ ] Unify agent architecture and implement agents with hierarchy (General = > Split (Traffic, Learning)).
• [ ] Create a general controller interface.
• [ ] Ground-Up rebuild of the communication buffer.