code_data: replace train.parquet with LCB-wrapped version; drop train-wrapped.parquet

Add LCB-wrapped-taco train variant (train-wrapped.parquet) to fix stdin-hang rollout stalls

Reorganize: drop -boxed/-buggy/-fixed variants, add metadata column for slime030 dispatch

Match upstream rllm-org/rllm:deepcoder — only wrap LCB rows with system prompt, leave primeintellect/taco/codeforces raw (byte-identical to source HF agentica-org/DeepCoder-Preview-Dataset)

Re-prep code data with correct per-source GT format (list for lcb/primeintellect/codeforces, dict for taco) and large_string schema

Fix code train parquet schema

Fix code train ground truth formats

Add deduplicated math + code data with system instructions