Add FastAPI inference service with Docker deployment

[bchadburn]

Summary

• Adds api/ package with FastAPI service (main.py), Pydantic schemas (models.py), and a singleton Recommender class that encapsulates the two-tower + LightGBM inference path
• Dockerfile (multi-stage, python:3.11-slim) and docker-compose.yml for one-command deployment
• requirements-api.txt with fastapi and uvicorn pinned separately from ML deps
• README "Running the API" section with endpoints table and example curl request

Test plan

• docker compose up — service starts, /health returns {"status": "ok"}
• POST /recommend with user_id=0, top_k=5 returns 5 items with valid fields
• POST /recommend with user_id=9999 returns 422 with error message
• Interactive docs load at http://localhost:8000/docs

Closes #1