Reject a repeated column name in Dataset.remove_columns

[LeSingh1]

The bug

Dataset.remove_columns does not deduplicate its argument, so a repeated name deletes the same feature twice:

d = Dataset.from_dict({"a": [1, 2], "b": ["x", "y"]})
d.remove_columns(["a", "a"])
# KeyError: 'a'

The second del dataset._info.features[column_name] raises a bare KeyError, with nothing pointing at the repeat. A duplicate is easy to produce when the list is built programmatically, and the error does not say what was wrong.

The fix

Reject the repeat, in the style of the missing-column check that already sits directly above it.

Scope

I originally had this covering select_columns too, where a repeat produced a Dataset whose column_names was ['a', 'a'] while its features had one entry. @AbdullahRasheed45's #8368 already fixes that, on both Dataset and IterableDataset, so I have dropped that half — I only spotted #8368 after opening this. What is left is Dataset.remove_columns, which #8368 does not touch, so the two do not overlap.

I chose rejecting over silently deduplicating because quietly changing the result seemed worse than saying what is wrong, and it matches how #8368 handles the sibling method. Happy to make it a dedup instead if you would rather.

Verification

The new test fails on main and passes with the change, for both in_memory values. Comparing full tests/test_arrow_dataset.py runs with and without the change shows no test that fails only with it. One failure is present either way, test_dataset_to_iterable_dataset, which needs PyTorch and is not installed in my venv.