Create shared openai-compatible inference adapter

[russellb]

There are multiple open PRs looking to add inference adapters for services that offer OpenAI-compatible APIs.

• databricks - add databricks provider #83
• sambanova - Add sambanova provider #136
• runpod - Add Runpod Provider #157
• vllm - Add vLLM inference provider for OpenAI compatible vLLM server #178

These adapters all use the OpenAI client and have the same config options (a URL and API key).

Options for moving forward:

1. Create a single shared adapter for OpenAI-compatible API endpoints.
2. If some cases require additional unique configuration, most of the code could be shared.
3. Create code to be shared by each adapter that uses an OpenAI-compatible API.

[ashwinb]

I support creating a shared code but not a shared adapter. I think implementation quality in terms of faithfulness of precise token-level behavior (on the backend) varies so it is best if we keep a top-level fork but share underlying code where it makes sense. I also plan to put together a set of tests which test tool calling behavior and other corner-case behavior we know can be problematic.

Please do hold off on doing this work for a couple days. There's some impending refactors incoming. I will be merging some PRs in the meantime to avoid duplicate work on behalf of the providers.

[terrytangyuan]

@ashwinb I have put together some work on this before noticing this issue and your comment: #195

Is it possible to merge that effort and then move them to reusable code in a separate PR (or as part of your incoming impending refactors) if a shared adapter is not preferred? It's a working adapter. This will unblock some of our existing tasks and avoid duplicating work.