Fix OOM: CPU-only PyTorch + fully lazy imports

Fix OOM: lazy-load ML models, rename to VectorMind

feat: add frontend and backend code for multimodal RAG knowledge assistant