Weight training (MoE)

Queue LOBO expert training and ridge manifest fitting on the prediction-worker GPU queue. Requires DB batches with daily weights and features; artifacts are written under WEIGHT_MODEL_ROOT on the prediction service.