Evaluation on LiveBench-Coding

[wasiahmad]

In this PR, we are adding evaluation support for LiveBench-Coding through nemo-skills. Primary changes are:

• Dataset download through nemo-skills/dataset/livebench-coding
• Evaluation logic implemented at nemo_skills/evaluation/evaluator/code.py

[wasiahmad]

Evaluation results based on Qwen2.5-Coder-32B-Instruct (paper (Table 2, page 18) reporting 56.8 pass@1).

Our evaluation gives (greedy decoding but still ran 10 times):

"pass@1[avg-of-10]": {
    "num_entries": 128,
    "avg_tokens": 282,
    "gen_seconds": 897,
    "accuracy": 54.843750000000014,
    "accuracy_statistics": {
        "avg": 0.5484375,
        "std_dev_across_runs": 0.009603692143013424,
        "avg_sample_std_dev": 0.01391647534032558,
        "std_err_across_runs": 0.003036954111898594
    }
}

[Kipok]

please update docs. Also Wasi, please start using -s for your commits so that DCO passes. At some point we will enforce it and not merge any PRs for which it doesn't pass