Stop restating quantized, and drop the unused default

Correct the published metadata

PP-OCRv6: XNNPACK detect is now static int8

Re-export under the get_model_schema contract; restructure to the MODEL_SPEC layout