Add support for selecting few-shot examples by id

[ErenAta16]

Custom tasks currently choose few-shot examples through few_shots_select (sequential, random, or balanced), which always picks from the pool programmatically. There was no way to pin a task to a specific, reproducible set of few-shot examples the way the LM-Evaluation-Harness id_sampler does, which matters when comparing against a published harness run or when a paper specifies exact few-shot examples by id.

LightevalTaskConfig gains two new fields:

• few_shots_id_column: name of a dataset column to use as a stable identifier for few-shot rows. When set, this value becomes the id of the corresponding few-shot Doc instead of the row's position in the split. Evaluation docs are unaffected and keep using the row position, so this is opt-in and backward compatible.
• few_shots_id_list: an explicit list of ids (matching few_shots_id_column, or the row position if it is not set) to use as few-shot examples, in the given order, instead of applying few_shots_select.

FewShotSampler picks up few_shots_id_list ahead of the existing sorting strategies and raises a clear error if any requested id is missing from the few-shot pool, rather than failing silently or picking something else instead.

Fixes #634