[FT] supporting train-time vs test-time metrics

## Issue encountered

I have reward functions where:

- At train time: stricter behavior, aka it's harder to get "correct" (1.0 reward)
- At test time: relaxed reward behavior

This is analogous to:

- Train time: high school students only being allowed non-graphing calculators during tests
- Test time: workforce lets employees use computers at will

From https://huggingface.co/docs/lighteval/en/quicktour#basic-usage:

> The syntax for the task specification might be a bit hard to grasp at first. The format is as follows:
> ```none
> {suite}|{task}|{num_few_shot}|{0 for strict `num_few_shots`, or 1 to allow a truncation if context size is too small}
> ```

This syntax well-supports few-shot behaviors, but does not support train-time vs test-time differences.

## Solution/Feature

I have a few ideas:

- Generalizing `num_few_shot` (third entry) to be arbitrary metadata
- Making a new CLI arg such as `--loose` or `--strict`

## Possible alternatives

Duplicating metrics by name, so one goes from `reward_function_x` to `reward_function_x_train` + `reward_function_x_test`. This feels hacky though, it's not using the framework to solve problems for me.