Add chat categorization dataset (#98)

This task is meant to test a couple things:
1. Classification -> both on common things where it is expected to
perform well (e.g., sentiment, toxicity -> which currently is always 0)
2. Structured json output -> the schema is nested, which confused some
of the smaller 7b models i tested out but works fine for llama 32b code
instruct (and OAI/anthropic)


Includes a couple common things like enums.