Upload retraining_results_v3.json with huggingface_hub

fix: declare train+test splits in dataset config

replace train split with parquet (55,014 rows, seed=42)

add test split (15%, seed=42, stratified on has_bias) — 9,709 rows

Upload JuaKazi Swahili ground truth v5 — 64,723 rows

initial commit