Add REPO_STRUCTURE.md documenting repository organizational architecture

[Copilot]

Comprehensive architecture documentation for the DARE repository, covering the training subsystem (verl/), evaluation subsystem (opencompass/), model definitions, and their interconnections.

• Top-level layout: Directory purposes and relationships at a glance
• verl/ deep dive: Trainer hierarchy (SFT/RL/DPO), worker roles (Actor/Critic/Rollout/Reward), inference backends (Fast-dLLM, lmdeploy, SGLang), sharding strategies, and utility organization
• opencompass/ integration: Model adapters for each dLLM family, evaluation config structure, and per-model example directories
• models/: HuggingFace-compatible model definitions with modified attention backends per model family
• recipe/ and scripts/: One-click training scripts organized by model×algorithm, plus eval/preprocessing/checkpoint-conversion utilities
• Architecture diagrams: ASCII system diagram, training pipeline flow, model type taxonomy (MDLM vs BDLM), and RL algorithm support matrix across model families
• Design notes: Dual-environment isolation, model-algorithm decoupling pattern ({model}_dp_actor_{algo}.py), FSDP + Ray orchestration

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.