Natural Language to SQL GRPO Example

**Is your feature request related to a problem? Please describe.**

Tunix currently lacks an example showing how to use GRPO with an external execution environment (e.g., a SQL database), making it unclear how to define structured rewards beyond text only tasks.

**Describe the solution you'd like**

I propose adding a small natural language to SQL GRPO example where the model generates SQL, the SQL is executed against a small SQLite database, and the reward reflects execution success and result correctness. If maintainers are open to this addition, I can implement a full example.

**Additional context**

This would give users a clear reference for integrating Tunix with external tools and for building verifiable RL tasks beyond GSM8K.

**Checklist**

- [x] I have searched the existing issues for similar feature requests.
- [x] This is not a support question (please use the "bug template" for that).