llm client base class and interface

[rlucas7]

This task is done when a common interface is implemented for llm clients.
And that the existing code for gemini and openi (from this PR) is refactored to use the common interface.

One of the issues is that different llm providers have different results sets and different methods. Eg. gemini-1.5-pro has a bunch of harm probabilities whereas openai does not.

Also the key/value names may differ slightly for some of the things that are supported across clients.

llm client class need:

• a method to generate relevances
• a method to parse the output into a set of str, str key/vals
• a method to get credentials

additional items likely to be desireable:

• handling of retry logic
• handling of prompts (for now keep the prompt orthogonal to client)