Apply the indices mapping in Dataset.unique

[LeSingh1]

Dataset.unique only takes the indices mapping into account when it has a different number of rows than the underlying table:

if self._indices is not None and self._indices.num_rows != self._data.num_rows:

An indices mapping can also repeat some rows and drop others while keeping the same length. In that case unique reads the raw table and returns values that are not in the dataset:

ds = Dataset.from_dict({"col_1": ["a", "b", "c"]}).select([0, 0, 1])
ds["col_1"]         # ['a', 'a', 'b']
ds.unique("col_1")  # ['a', 'b', 'c']  <- 'c' is not in the dataset

class_encode_column builds its ClassLabel names from unique, so it inherits the bug and creates a class for rows that were dropped:

ds.class_encode_column("col_1").features["col_1"].names  # ['a', 'b', 'c']

This is reachable from ordinary use — select with repeats (bootstrap resampling), and any indices mapping whose length happens to match the table.

This applies the indices mapping whenever there is one. Only the requested column is materialized, via the existing query_table(..., key=column, indices=self._indices) path, so it deliberately avoids falling back to flatten_indices() — that would make shuffle().unique() rewrite the whole dataset.

Added test_unique_with_indices_mapping_of_same_length and test_class_encode_column_with_indices_mapping_of_same_length; both fail on main with assert ['a', 'b', 'c'] == ['a', 'b'] and pass with the change. tests/test_arrow_dataset.py, tests/test_dataset_dict.py and tests/test_search.py pass (470 passed, 38 skipped).