Multi-worker evaluation fails: CUDA multiprocessing error

Hi, I used `--num-workers > 1` in evaluation, and it caused the evaluation to fail with:

```
Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method
```

I ran the following command to start evaluation:

```
olmes \
--model allenai/OLMo-2-0425-1B \
--revision stage1-step1907359-tokens4001B \
--model-args '{"trust_remote_code": true}' \
--task olmo3:base:math \
--batch-size 32 \
--num-workers 8 \
--output-dir outputs/OLMo-2-0425-1B-stage1-real_olmo3:base:math_01 \
--limit 0.1
```

Does your evaluation code support multi-worker evaluation?