Fix balanced few-shot sampling: falsy label truncation and global RNG use

[ErenAta16]

Closes #1309.

What

FewShotSampler._init_fewshot_sampling_balanced (src/lighteval/tasks/prompt_manager.py) had two issues in the same function.

1. A falsy class label silently truncates the balanced sample. The label-cycle guard was if not next_label: break. labels_iterable is a cycle, which never yields None for a non-empty pool, so the guard was meant for the empty-pool case, but if not next_label also fires on a valid but falsy label (integer 0, empty string, False). Labels come from instance.fewshot_sorting_class or as_list(instance.get_golds())[0], so any task with a falsy label had its balanced selection cut short, down to zero examples if that label sorted first.

2. Balanced sampling used the global random module. Unlike the sibling _init_fewshot_sampling_random (local random.Random(variance_seed)), the balanced path called random.seed/random.shuffle/random.randrange on the global module, which is non-reproducible if other code touches the global RNG and leaks state into the rest of the program.

Reproduction (before)

from lighteval.tasks.prompt_manager import FewShotSampler
from lighteval.tasks.requests import Doc

class FakeTask:
    fewshot_selection = "balanced"; fewshot_split = "test"
    def __init__(self, docs): self._docs = docs
    def fewshot_docs(self): return self._docs

docs  = [Doc(query=f"q{i}", choices=["", "x"], gold_index=0) for i in range(20)]  # label ""
docs += [Doc(query=f"p{i}", choices=["", "x"], gold_index=1) for i in range(20)]  # label "x"
print(len(FewShotSampler(FakeTask(docs)).sample_fewshot_examples(10, variance_seed=0)))  # 0, expected 10

Fix

1. if not next_label -> if next_label is None (a cycle only yields None when sorted_labels is empty; falsy-but-valid labels no longer truncate).
2. Use a local rnd = random.Random(variance_seed) and its .shuffle/.randrange, mirroring _init_fewshot_sampling_random.

Tests

Added two regression tests to tests/unit/prompt/test_prompt_manager.py:

• test_balanced_fewshot_falsy_label_not_truncated: a task with an empty-string-gold class returns the full num_fewshot with both classes represented.
• test_balanced_fewshot_does_not_disturb_global_rng: a balanced sampling call leaves the global random stream untouched.

RED/GREEN verified: both fail on main and pass with the fix. test_prompt_manager.py is green (5 passed), ruff check clean on both files.