Batching

[Techbert08]

This does three things (I love threes):

1. Update default config to explicitly use 4.1 mini for scoring. This makes it more clear what's in use.
2. Extending the LLM interface to have a "get_completions" in parallel to "get_completion". The default providers just route "get_completions" to parallel-call "get_completion". But the Pi one instead uses first-class support for parallel calls. There's some yuckiness around lists of kwargs vs kwargs.
3. I added "ask_llm_parallel" to expose this functionality to the ranker.

Now it works with a local Pi modelserver and with 4.1 mini. My plan is to merge this (assuming it looks good), then finish deploying the updated Pi modelserver to Azure.