feat: VLLMModel has option to spin up local vllm

[bxyu-nvidia]

Use cases, pain points, and background Right now users have to point to some external model endpoint. We should enable users to spin up their own.

Description: Add an option to spin up a local vllm server in VLLMModel.

Design: Probably just responses_api_models/vllm_model app.py and requirements.txt

Out of scope:

Acceptance Criteria:

• [ ] Implementation of the above, along with training curves using the spun up vllm endpoint as a judge or something.