Add openai_eval type to delegate evals to OpenAI APIs

[krisztianfekete]

This PR adds support for OpenAI's Evals API as an evaluator backend in agentevals. Instead of running grading logic locally, it delegates to OpenAI's hosted evaluation infrastructure. The first supported grader is TextSimilarityGrader (fuzzy_match, BLEU, ROUGE, cosine, etc.), but the design is set up so adding more grader types later is straightforward.

We introduce a new evaluator type, openai_eval.

You can configure it like this:

evaluators:
  - name: response_similarity
    type: openai_eval
    threshold: 0.8
    grader:
      type: text_similarity
      evaluation_metric: bleu

Under the hood, the backend does a create-eval, create-run, poll, collect-results, cleanup cycle against the OpenAI API. To make it work for pre-existing agent traces (because we just want grading) this puts both the actual and expected text into the item namespace with include_sample_schema: False, so OpenAI never tries to generate model outputs.