Support OpenJury & Evalchemy in evals

Add support in https://github.com/OpenEuroLLM/multi-cluster-eval to run evaluations from https://github.com/mlfoundations/evalchemy/

In particular, we should make sure that at least Alpaca-Eval with a local model judge served with vllm works, e.g. to make sure that we have a good workflow to run this:

python -m eval.eval \
    --model vllm \
    --tasks alpaca_eval \
    --model_args "pretrained=meta-llama/Meta-Llama-3-8B-Instruct" \
    --batch_size 16 \
    --output_path logs

(we may want to start by just exposing a clear way to do this with the output paths specified from environment variables rather than having a CLI to schedule such evals)

Start by reviewing the multi-cluster-eval repository and the evalchemy invocation shown in the issue, including how evaluation output paths are handled. Verify the Alpaca-Eval workflow with a local model judge served by vLLM; done means the evaluation can run with output paths supplied through environment variables or a clearly documented scheduling interface.