Zephyr : Stitching Heterogeneous Training Data with Normalizing Flows for Photometric Redshift Inference