Add Belief retrieval benchmark samples (train/validation/test, parquet+jsonl.gz)

Align retrieval organization: keep support_set under dataset; benchmark/Belief/retrieval reserved for benchmark samples