[recipe] Add GRPO training recipe for cosyvoice2 llm

[yuekaizhang]

See #1463

Also, added a script to convert the RL trained model back into CosyVoice Repo compatiable format.

[ScottishFold007]

这个貌似不太稳，还是试试DPO吧

Cosyvoice DPO训练笔记：https://github.com/ScottishFold007/Cosyvoice_DPO_NOTES

点开MD文档看详情：

[xiaoyangnihao]

@yuekaizhang 在token2wav_asr_server中，保存token2wav之后的音频，每个音频均是20s，然后末尾是大段的静音段或者噪音，感觉训练中模型似乎没有很好的学习到eos