ML on the digital twin: export, train pilots, and import predictions back into the app

Part 9 of 12 — mix export.ml CSVs from TimescaleDB, optional Elixir pilots (Scholar/Axon), JSONL import into ml_predictions, and MlPredictionsLive.