feat: add zero_division parameter to F1 metric

[YousefZahran1]

What

precision and recall both accept a zero_division parameter that controls what value is returned when a label has no predicted or true samples. f1 does not, which causes:

1. A misleading user experience — sklearn's UndefinedMetricWarning for F1 explicitly tells users to pass zero_division, but the evaluate wrapper silently drops it with a TypeError.
2. An inconsistency within the same library (precision and recall have it; f1 doesn't).

Reproduce the bug

import evaluate
f1 = evaluate.load("f1")
# sklearn's warning on this call says: "Use zero_division parameter to control this behavior."
f1.compute(predictions=[0,0,0,0,0], references=[0,1,0,1,2],
           average=None, labels=[0,1,2,3], zero_division=0)
# TypeError: F1._compute() got an unexpected keyword argument 'zero_division'

Fix

Add zero_division="warn" to F1._compute() (matching the default in precision and recall) and pass it through to sklearn.metrics.f1_score.

Changes

• metrics/f1/f1.py: add zero_division argument to _compute(), document it in _KWARGS_DESCRIPTION, and add Example 6 showing the parameter in action.

Testing

f1 = evaluate.load("f1")

# zero_division=0: undefined labels get 0
result = f1.compute(predictions=[0,0,0,0,0], references=[0,1,0,1,2],
                    average=None, labels=[0,1,2,3], zero_division=0)
# {'f1': [0.57, 0.0, 0.0, 0.0]}

# zero_division=1: undefined labels get 1
result = f1.compute(predictions=[0,0,0,0,0], references=[0,1,0,1,2],
                    average=None, labels=[0,1,2,3], zero_division=1)
# {'f1': [0.57, 0.0, 0.0, 1.0]}

# default (no zero_division): backward-compatible — still warns and returns 0
result = f1.compute(predictions=[0,0,0,0,0], references=[0,1,0,1,2],
                    average=None, labels=[0,1,2,3])
# UndefinedMetricWarning raised, {'f1': [0.57, 0.0, 0.0, 0.0]}

Fixes #699