fix(metrics): let _mean ignore unmeasured scores like its siblings

[sclfcz]

Fixes #4509

What

_mean documents returning None when there is nothing to average, but unlike its neighbours _stdev and _pstdev it does not drop None values before averaging, so a score list with an unmeasured entry raised instead of being averaged:

call	before	after
_mean([0.9, None, 0.8])	TypeError: can't convert type 'NoneType' to numerator/denominator	0.85
_mean([None, None])	TypeError	None
_mean([])	None	unchanged
_mean([1.0, 2.0])	1.5	unchanged

_stdev([0.9, None, 0.8]) → 0.071 and _pstdev → 0.05 already filter those values, and their type hint is List[Optional[float]]; _mean now matches them (hint included).

Testing

Added test_mean_ignores_unmeasured_scores to test_unstructured/metrics/test_utils.py covering the four rows above.

pytest test_unstructured/metrics/test_utils.py → 6 passed; reverting the filter fails exactly the new test and leaves the other five passing.

[chrikrah]

@sclfcz I would merge this. The note below is non-blocking.

non-blocking: real callers aggregate a pandas column, so a missing score arrives as NaN, not None, and _stdev raises before _mean is reached. get_mean_grouping with one blank cct-accuracy cell fails identically before and after this change:

$ python probe_group.py   # 3 rows, cct-accuracy [0.9, None, 0.8], group_by="doctype"; torch and unstructured_inference stubbed
main ddf4453:  raised: ValueError inf or nan encountered in data
PR   43867d0:  raised: ValueError inf or nan encountered in data
# Python 3.13.15, pandas 2.3.3

Would you like pd.notna in _mean, _stdev and _pstdev as part of this change, or kept for a follow-up?