Improve NarrativeQA metrics and prompt structure

## Describe the bug
`narrative_qa_helm`:
- The prompt function signature and `Doc` creation had a nested choices issue.
- The benchmark was configured with `Metrics.exact_match`, which is inappropriate for long-form generative reading comprehension. This led to silent failures (0.0 score), even when using fairly powerful models like `DeepSeek-V3.2`, whereas switching to other metrics work better. Although providing some few shot examples slightly improved the model's performance on exact match metric, but still I think there are better metrics to use

## To Reproduce
```python
task = "narrativeqa|0"

pipeline = Pipeline(
tasks=task,
pipeline_parameters=pipeline_params,
evaluation_tracker=evaluation_tracker,
model_config=model_config,
)

pipeline.evaluate()
pipeline.save_and_push_results()
pipeline.show_results()

```

## Expected behavior
- The benchmark should use `rougeL` and `f1_score` metrics.
- The prompt function should provide a flat list of references in the `Doc.choices` field.

## Version info
- OS: mac
- Lighteval version: main (local development)