Refactor LLM architecture to SOLID principles, add Gemini Judge, and self-reflective RAGAS logic

security: Remove hardcoded GLM API key from source files (use environment variable GLM_4_7_API_KEY)

feat: Implement GLM-4.7-Flash RAGAS Judge and LLM fallback (Qwen -> GLM-4.7-Flash)

Fix file upload UI refresh and add Metadata Filters UI

Initial commit: Multimodal RAG Pipeline V3.0 with Fallback Logic