feat: revert backend to use sentence-transformers (all-MiniLM-L6-v2) for better embeddings

feat: add Dockerfile for Hugging Face Spaces deployment

refactor: replace memory-heavy PyTorch with fastembed (ONNX) to fix OOM

perf: limit in-memory corpus to 5k docs to prevent OOM on 512MB free tier

fix: precompute dense embeddings to prevent OOM on render

Initial commit