Single controller

[keshavb96]

This PR demonstrates how to use the building blocks in the bridge codebase to implement orchestration through a single controller. In totality we now have:

1. Decoupling of processes that do weight updates / training, prompt dispatching and rollout generation (vLLM). See example in examples/decoupled_synchronous.
2. Async RL support, see example in examples/decoupled_asynchronous
3. Single-controller (sync and async) example demonstrating how to orchestrate the RL training loop + inference offloading with a single-controller using the building blocks of the bridge.