Add `logprobs` parameter

Looks like it's now in OpenAI's chat completion endpoint so we will definitely need to support that.

Optionally, for tooling maybe allow a minimum prob threshold to use the tool (if not at a threshold, ask for clarification)