Horizontal merge insert should not rewrite join columns

When we perform a horizontal merge_insert we rewrite all columns that are modified. The join column cannot be modified and should not have to be rewritten.

However, we are rewriting the join column, which is slightly wasteful.

import lance
import pyarrow as pa
import shutil

shutil.rmtree("/tmp/test_dataset", ignore_errors=True)

ds = lance.write_dataset(pa.table({"id": range(100), "value": range(100), "other": range(100)}), "/tmp/test_dataset")
ds.merge_insert("id").when_matched_update_all().when_not_matched_insert_all().execute(pa.table({"id": range(50, 150), "value": range(50, 150)}))

frag = ds.get_fragments()[0]
data_file = frag.data_files()[0]

print(f"Fields={data_file.fields}") # This is actually [-2, -2, 2]
assert data_file.fields == [0, -2, 2]

Start with the merge_insert entry point shown in the Python reproduction and trace its Rust implementation to where updated fields are selected for horizontal merges. Use the supplied dataset and assertion to inspect the resulting data_file.fields. Done means the join column remains unchanged while modified columns are rewritten, with the reproduction passing.