Core Data Models

[vidhyav]

Rithesh's pr commandeered: #157

[allenwang28]

Can we do this incrementally? Let's start a PR that integrates completion and experience, with Experience renamed to Episode (consistent with what we have now)

I want that, plus the ability to convert from e.g. vLLM native outputs to our Forge native outputs.

An example from GRPO main: https://github.com/meta-pytorch/forge/blob/main/apps/grpo/main.py#L311-L313, we should be able to convert from the vLLM request to a forge.data_types.CompletionOutput, then create an Episode from only forge native data types. Ideally like:

responses = await policy.generate.choose(prompt)
completions = forge.data_type.CompletionOutput.from_vllm(responses)

etc.

[vidhyav]

Huh this was a set of stacked diffs and it is painful to commandeer AND modify. I am going to simply restart a new diff. SIGH.