cast_column to ClassLabel silently accepts out-of-range label indices

[shoutoutuoadi325]

Describe the bug

Dataset.cast_column / Dataset.cast to a ClassLabel column silently keeps label indices that are out of range for the target number of classes. The corruption only shows up later — e.g. on the first int2str call — or silently trains on wrong labels:

from datasets import Dataset, ClassLabel

d = Dataset.from_dict({"label": [5, 1]})
casted = d.cast_column("label", ClassLabel(names=["neg", "pos", "oth"]))  # no error, num_classes=3
print(casted["label"])                    # [5, 1]
casted.features["label"].int2str(5)       # ValueError: Invalid integer class label 5

The write path does validate the same data — Dataset.from_dict(..., features=Features({"label": ClassLabel(names=[...])})) raises ValueError: Class label 5 greater than configured num_classes 3 — so the cast path bypasses a check the writer performs.

Cause: table_cast in src/datasets/table.py only runs the feature-level casts when the arrow schema changes, and pa.Schema.__eq__ ignores metadata. Casting int64 storage to ClassLabel leaves the arrow schema unchanged and only alters the metadata, so it takes the replace_schema_metadata branch and ClassLabel.cast_storage, which does the range check, never runs.

Expected behavior

cast_column/cast should raise the same ValueError as the write path when a label index is >= num_classes.