[Ready] cosyvoice llm rl for TTS

[yuekaizhang]

This PR supports using verl to conduct RL training experiments on cosyvoice2 llm models.

Specifically, we conducted GRPO experiments and got below results:

Model	Seed-TTS test_zh CER	Cosyvoice3 zero_shot_zh	Comment
Official CosyVoice2 LLM	1.45 %	4.08%	See the paper
SFT (initialized from Qwen2-0.5B-Instruct)	1.81 %	4.83%	See PR #1887
GRPO (this work, trained on AIShell-3)	1.06 %	4.03%