Fix simultaneous label remapping in mean_iou

[feng1201]

label_map currently runs before converting decoded images to NumPy arrays. Consequently, MeanIoU.compute raises TypeError: 'PngImageFile' object does not support item assignment for non-empty mappings. When the helper receives a NumPy array directly, successive assignments instead cascade and modify the caller's labels: {0: 1, 1: 0} turns [[0, 1]] into [[0, 0]] instead of swapping the classes.

Convert inputs to independent arrays first, then match every mapping against an unchanged copy of the original labels. Document the simultaneous mapping behavior. Label reduction and ignored-pixel handling remain in their existing order.

Reproduction from the repository root:

import evaluate
import numpy as np

metric = evaluate.load("./metrics/mean_iou")
result = metric.compute(
    predictions=[np.array([[1, 0]], dtype=np.uint8)],
    references=[np.array([[0, 1]], dtype=np.uint8)],
    num_labels=2, ignore_index=255, label_map={0: 1, 1: 0},
)
assert result["mean_iou"] == 1.0

Validation on Python 3.11.15, Datasets 2.21.0, NumPy 1.26.4 and Pillow 12.3.0:

• New regression suite on unchanged source: 18 failed, 4 passed.
• python -m pytest tests/test_mean_iou.py tests/test_metric_common.py -k mean_iou -q --tb=short: 23 passed, 1 remote test skipped by the repository's default settings.
• make quality: passed across the repository.

The regressions cover public load(...).compute(...) calls with NumPy/PIL inputs, binary swaps, three-class cycles, mapping order, read-only arrays, caller input preservation, and label reduction/ignored pixels. Other framework-dependent metrics and remote tests were not run. Prepared with Codex assistance; the reported commands were executed locally.