Load full H5 model via load_model; pin TF 2.20 to match training runtime

Fix build_model to exactly match train.py architecture and preprocessing

Fix missing layers import in build_model

Clean deployment: model loaded from HF Hub at runtime