Add LongBench V2 benchmark

[eshwarprasadS]

Adding LongBench to eval options,

Install extras with:

pip install instructlab-eval[longbench]

Uses VLLM backend for serving the model for generation

Runs like so:

evaluator = LongBenchEvaluator(
    model_path="path/to/model",
    num_gpus=N,
    output_file="path/to/results.json",
    eval_config={"batch_size": "auto"},
    vllm_config={"max_model_len": max_len}
)

results = evaluator.run()  # Returns LongBenchResult

Output json looks like so:

{
  "en_multidoc": 0.5424139838230786,
  "zh_multidoc": 0.24335639081098673,
  "en_singledoc": 0.4233139199560039,
  "zh_singledoc": 0.46157875457875464,
  "en_summ": 0.27244809337990245,
  "zh_summ": 0.1359562304911904,
  "en_fewshot": 0.45692449627485754,
  "zh_fewshot": 0.24416666666666667,
  "en_synthetic": 0.3799285714285714,
  "zh_synthetic": 0.4775,
  "code_avg": 0.30225,
  "overall_score": 0.3581670097645466
}

[RobotSail]

Thanks for the PR @eshwarprasadS !

The PR has all of the right ideas, there are just a few minor changes that you'll want to make which I've outlined in this review. Once we've addressed those, this should be good to merge

[RobotSail]

Why do we check if k != "overall_score"? We shouldn't have set this key yet