Add 10k RL puzzle split

Update SFT split with tile-correct rationales

Replace SFT data with 600 depth 8-15 puzzles

Upload data/eval-00000-of-00001.parquet with huggingface_hub

Upload data/sft-00000-of-00001.parquet with huggingface_hub