chains/benchmarks, other LLMs

[dovinmu]

(closing #28 in favor of this)

At minimum before we merge this branch:

• [x] generalize the llm.py op to route to different specific LLM sources
• [x] generic bigbench binary classification script

for the binary classification script:

• [x] accepts LLM ID and task name as input
• [x] calculates at least one accuracy metric
• [x] puts results somewhere useful (json file?)

these are nice to haves, basically they'd make it so we were using the bigbench tasks more as intended:

• [ ] (blocked) need to be able to extract multiple key/values from json
• [ ] (blocked) use "preferred_metric" from the task.json spec to decide metric

If #27 is resolved then we could compute multiple accuracy metrics. And we could also optionally explore having another script that runs the entire slate of tasks on a set of models.

[dovinmu]

I ended up writing a few more ops for this PR, some of which seem obvious (!csv-parse) and some of which I expect will change (the !metrics- ones). I'm especially interested in getting the changes to llm.py merged so we have a way of calling LLMs from different sources