feat: add TSP (Tree-like Self-Play) for Secure Code LLMs

[Easonnoway]

Summary

Add the Learn from Your Mistakes: Tree-like Self-Play for Secure Code LLMs project to the hub.

TSP is a training framework that improves code generation security by identifying CWE risk nodes in the generation process and using targeted self-play with DPO training to teach models to distinguish secure code paths from vulnerable alternatives.

This PR includes the core TSP pipeline:

• Step 0: CWE risk node annotation via GPT-4o
• Step 1: vLLM code generation at vulnerability decision points
• Step 2: DPO preference pair creation and format conversion
• Step 3: DPO training via LLaMA-Factory (with configs for CodeLlama-7B and Qwen2.5-Coder-7B)

Test plan

• Verify directory structure matches hub conventions (similar to IsYourPromptPoisoningCode)
• Confirm README format is consistent with existing project
• Review that no hardcoded secrets or absolute paths are present
• Check Python syntax for all scripts