Drop the family-level parquet dirs superseded by one subset per task

One subset per task

Rebuild on the benchmark's own answer protocol; columns id, query, response, scoring_context first

Publish the hub-native mirror: unified query/ground_truth/response/subtask/scoring_context columns beside every upstream column

initial commit