Adds Math-500 and AIME24 evals

[edbeeching]

Evaluating models (internal)

For small models use --data_parallel=$NUM_GPUS, for large models shard with --tensor_parallel=$NUM_GPUS
Example for evaluating deepseek-ai/DeepSeek-R1-Distill-Qwen-7B

NUM_GPUS=1
MODEL=deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
MODEL_ARGS="pretrained=$MODEL_ID,dtype=bfloat16,data_parallel=$NUM_GPUS,max_model_length=4096,gpu_memory_utilisation=0.8"
TASK=aime24 # or math
OUTPUT_DIR=evals/$MODEL

lighteval $MODEL_ARGS $TASK --use-chat-template --custom-tasks src/open_r1/eval/$TASK.py --output-dir $OUTPUT_DIR --system-prompt="Please reason step by step, and put your final answer within \boxed{}."

[lewtun]

LGTM!