lighteval

1958f0d2 - Regenerate SmolLM2 slow-test references for transformers 5 / vllm 0.30

Regenerate SmolLM2 slow-test references for transformers 5 / vllm 0.30 Regenerated on the CI T4 runner inside the pytorch 2.13 container. Only arc:challenge acc moves beyond tolerance (0.4 -> 0.3, one of 10 samples); truthfulqa_mc2 shifts by ~0.001. Also removes the temporary regeneration steps from the workflow. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>