[C++] Acero/dataset relies on ExecBatch::ToRecordBatch truncating excess columns

As found while working on ARROW-18004: the dataset scanner and the Acero engine rely on `ExecBatch::ToRecordBatch` returning successfully when the given schema has fewer fields than the ExecBatch has columns.

This apparently allows to implicitly drop the dataset-added columns (`kAugmentedFields` in `arrow/dataset/scanner.cc`) from a scan's final result.

However, it seems wrong and brittle to do this implicitly at the `ExecBatch::ToRecordBatch` level (hiding potential errors). Instead, it should probably be done explicitly inside Acero/dataset.

**Reporter**: [Antoine Pitrou](https://issues.apache.org/jira/browse/ARROW-18037) / @pitrou
#### Related issues:
- [[C++] ExecBatch conversion to RecordBatch may go out of bounds](https://github.com/apache/arrow/issues/33208) (is related to)

_{**Note**: *This issue was originally created as [ARROW-18037](https://issues.apache.org/jira/browse/ARROW-18037). Please see the [migration documentation](https://github.com/apache/arrow/issues/14542) for further details.*}

Start by reading ExecBatch::ToRecordBatch and the dataset scanner code at arrow/dataset/scanner.cc, then trace the Acero/dataset paths that depend on schemas with fewer fields. Compare the behavior with related issue #33208. Done means column dropping is explicit in Acero or dataset rather than implicitly accepted by ExecBatch::ToRecordBatch, with coverage for the affected scan or execution path.

Written by the indexing model from the issue text.