supervised fine-tuning

Parallel-SFT aims to make code RL transfer better

Parallel-SFT uses equivalent code across languages to improve zero-shot transfer in code RL, especially when moving to lower-resource programming languages.