Fix flatten_dict crash/wrong key for bare numpy array values

[Kymi808]

Summary

In flatten_dict (src/lighteval/utils/utils.py), the recursive helper's np.ndarray branch builds its key using i:

elif isinstance(v, np.ndarray):
    into[prefix + k + sep + str(i)] = v.tolist()

But i is the loop variable from the preceding list/tuple branch (for i, vv in enumerate(v)). For a standalone ndarray value it is never bound, so the call raises UnboundLocalError. If a list/tuple key was processed earlier in the same dict, the stale i leaks in and produces a bogus indexed key.

flatten_dict({"a": np.array([1, 2, 3])})
# UnboundLocalError: cannot access local variable 'i'

flatten_dict({"lst": [10, 20, 30], "arr": np.array([7, 8, 9])})
# {'lst/0': 10, 'lst/1': 20, 'lst/2': 30, 'arr/2': [7, 8, 9]}   <- bogus '/2'

flatten_dict feeds obj_to_markdown, which evaluation_tracker uses to log config/task details to TensorBoard, so an ndarray-valued config entry crashes logging.

Fix

A bare ndarray maps to a single value, so key it as prefix + k, identical to the scalar branch just below it:

elif isinstance(v, np.ndarray):
    into[prefix + k] = v.tolist()

The list/tuple-of-ndarrays path (which legitimately uses i) is unchanged.

Test plan

• Added TestFlattenDict covering the unbound case, the stale-index case, and a regression guard for list-of-ndarrays. The first two fail on main (UnboundLocalError / bogus arr/2 key) and pass with this change.
• pytest tests/unit/utils/test_utils.py → 9 passed.
• ruff check / ruff format --check clean on both files.