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

Add jsonl.gz exports for benchmark/Belief/profile (for Belief inference)

Add jsonl.gz exports for benchmark/Belief/direct (for Belief inference)

Canonicalize top-level benchmark/ hierarchy (Belief + Trade)