Rework for New StreamingLearnerInvestigator

[BenCarter44]

This PR will fulfill issue #28 .

Here below is a draft spec of the new StreamingLearnerInvestigator.

A user simply subclasses it, and can add their own logic (and their own ROSE learner) to the callbacks. This StreamingLearnerInvestigator simply provides the batch support and plumbing.

Will need to be tested and cleaned up....

(Tests right now aren't expected to work, as this is a draft spec)

[andre-merzky]

I know it is early, but quick question: why are we concerned about batch size on this level of abstraction? The Orbit layer is already batching requests and thus optimizing network latency impact. What is the purpose of additional higher level batching?

[BenCarter44]

Hello Andre,

I'm not as familiar with ORBIT's batching. Could you describe more?

I'm highly sure the batching here in StreamingLearnerInvestigator is different. The investigator still gets input one-by-one. The "batch" here is simply a group of investigator inputs sent to the active learner.

It's role is to simply answer: "instead of triggering an AL on every input, trigger it every X inputs, and group them up as you go".

Note that this batching does not affect inference, as inference runs on each input one-by-one to the investigator.