Fix IterableDataset.remove_columns when a single column name is passed

[LeSingh1]

IterableDataset.remove_columns accepts either a single column name or a list, but it only ever tests col in column_names. When a string is passed, that becomes a substring test, so any existing column whose name is contained in the one being removed is dropped from info.features:

ds = Dataset.from_dict({"label": [0, 1], "label_text": ["a", "b"]})
it = ds.to_iterable_dataset().remove_columns("label_text")

it.column_names  # []  -> should be ['label']
it.features      # {}
next(iter(it))   # {'label': 0}  -> the column is still being yielded

features and column_names then disagree with the actual examples, which breaks anything that reads the schema afterwards — select_columns, cast, or writing to parquet. The map-style Dataset.remove_columns("label_text") correctly returns ['label'], so the two paths disagree.

The existing test_iterable_dataset_remove_columns passes only incidentally: it removes "id" and no sibling column name contains "id".

This normalizes a string argument to a list first, which is what select_columns in the same class already does.

Added coverage for the single-string case with a sibling whose name is a substring; it fails on main with assert [] == ['label'] and passes with the change. tests/test_iterable_dataset.py and tests/test_dataset_dict.py pass (531 passed, 30 skipped; the 2 remaining failures are pre-existing network-dependent tests unrelated to this change).