Refactor recommender into train/serve pipeline with ALS recommendations and Docker uv setup

[rasadov]

Summary

• Rebuilds the Python recommender as three processes: Kafka sync (consumer), offline training (train), and gRPC serving (server), all via a single main.py entrypoint.
• Adds a recommendation pipeline using implicit ALS (personalized), item similarity (viewed products), and popularity fallback, with models saved to a shared artifacts volume.
• Introduces a pluggable file manager (local disk / S3-compatible bucket) for model artifact storage.
• Fixes integration with the rest of the stack: gRPC listens on port 8080, Kafka routing uses upstream topic names, and DB schema init is safe for concurrent container startup.
• Updates Docker to use uv + uv.lock, adds a scheduled recommender-train service, and wires artifact volumes across train and serve containers.