Align `queries` schema with the q2c chunk-retrieval benchmark family (drop stray `topic` column)
FreshStackChunkRetrieval follows the shared q2c chunk-retrieval / chunk-size-sensitivity benchmark format (QasperChunkRetrieval, the ConTEB *ChunkRetrieval sets, etc.), where every dataset's queries config uses the same schema:
original_query, query, answer, score, source_url, frag_start_char, frag_end_char, n_gold
FreshStack's queries parquet diverges from this: it carries an extra topic column that is not part of the shared schema and is not in this config's own declared features. That makes the config incompatible with the standard loader — load_dataset(repo, 'queries') fails while casting the parquet to the declared features (the extra column has no match), so any tool built on the shared format breaks at load time. (The documents-* configs also carry topic, but declare it consistently, so only queries is affected.)
This PR drops the stray topic column from queries so its schema matches the declared features and the rest of the benchmark family, making the config loadable by the standard pipeline. Row count is unchanged (672); no other column is touched.
If topic is intentional, the alternative is to declare it in the queries features so data and schema agree — happy to switch to that instead.