[new recipe] Add Cosyvoice TTS GRPO training recipe based on veRL.

[yuekaizhang]

This PR introduces a recipe 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 CER (%)⬇️	Comment
SFT (initialized from Qwen2-0.5B-Instruct)	1.81 %	4.83%	See PR #1887
GRPO (this project, trained on AIShell-3)	1.06 %	4.03%	See here

Features:

1. Using Pytriton based Sensevoice ASR sever to achieve fast reward calculation.
2. Using phoneme error rate (PER) as the reward metrics.
3. Support both pretrained cosyvoice2 llm and custom sft version of cosyvoice2

[gemini-code-assist]

Code Review

This pull request adds a new project, CosyVoice-TTS-GRPO, to the list of projects using verl in the main README.md. The change is straightforward, but I've pointed out a minor formatting inconsistency that should be addressed to maintain consistency with the rest of the document.

[eric-haibin-lin]

nice! do you plan to contribute the full recipe, or just the readme? BTW please resolve conflict with main branch. thanks!

[yuekaizhang]

nice! do you plan to contribute the full recipe, or just the readme? BTW please resolve conflict with main branch. thanks!

I'd love to contribute the recipe here. Let me update the recipe into this PR.

[CLAassistant]

All committers have signed the CLA.

[yuekaizhang]

@eric-haibin-lin Updated. Would you mind checking it again? Many thanks.