Stop restating quantized, and drop the unused default

Core ML: drop fp32, fp16 is equivalent in use

fix(v0.9.0): switch to size-first layout (s/{xnnpack,coreml}/ + x/{xnnpack,coreml}/)