Add agent manager

[vsoch]

This pull request adds an agent manager. This is how it works.

1. The user defines a plan. The plan has an overall goal (used by the manager) along with some number of steps. Each step is owned by an agent.
2. The manager loads and validates the plan, and creates a common context. The context is a modified dictionary object that is updated by each step, ensuring that we pass a common state between agents, and can return a consistent set of variables to the manager (typically a final result and return code).
3. Each agent internally has a goal - e.g., a successful docker build, or generation of a Kubernetes Job and running the job. Each agent can take a maximum number of attempts to get to that state, with the default being unset (keep trying forever).
4. If the max attempts is set for an agent, it will return early and give the state to the manager. The manager can look at it, and decide how to proceed. This means moving to give the context to a different agent. For example, if the build has a library issue discovered in the Kubernetes job run, we'd want to give that back to the build agent.
5. The manager also can take a maximum number of attempts for trying each step, external to the step.
6. In any case that the manager takes over, the manager first selects the agent to use, and then asks the agent for the prompt it would provide given the state. The manager is then given the entire context and asked to tweak or change the prompt to be more fitting to what needs to be done.

Right now I'm finishing up the basic orchestration, and I need to finish the implementation of the manager to ask for prompts. When I finish that, I'll work on an optimization agent, which will take the log from an application run, and work on primarily the entrypoint command (and manifest) to make it run better.