Bump LightEval to enable DP>1

[lewtun]

This PR bumps lighteval to enable DP>1 again since it is compatible with vllm v0.8.4. See: huggingface/lighteval#670 (comment)

Waiting for evals to run, but code can be reviewed as is.

Update: evals finished and the diff is up to a few percentage points, but most on AIME24 which is rather noisy due to small sample size:

Code snippet to run all evals:

#!/bin/bash

MODELS=(
"DeepSeek-R1-Distill-Qwen-1.5B"
"DeepSeek-R1-Distill-Qwen-7B"
"DeepSeek-R1-Distill-Qwen-14B"
"DeepSeek-R1-Distill-Qwen-32B"
"DeepSeek-R1-Distill-Llama-8B"
"DeepSeek-R1-Distill-Llama-70B"
)

for M in "${MODELS[@]}"; do
  echo "Running benchmark for model: $M"
  python scripts/run_benchmarks.py --model-id deepseek-ai/$M --benchmarks aime24 math_500 gpqa lcb
done

TODO

• Re-run evals to check score variance

[lewtun]

max_num_batched_tokens is no longer needed to be included, so I've removed it for simplicity