lighteval

1607dc10 - Adds multimodal support and MMMU pro (#675)

Adds multimodal support and MMMU pro (#675) ``` uv run lighteval accelerate "model_name=HuggingFaceTB/SmolVLM-Instruct" "lighteval|mmmu_pro|0|0" --use-chat-template --vision-model ``` --------- Co-authored-by: qubvel <qubvel@gmail.com>