Declare the int8 Core ML variant

Stop restating quantized, and drop the unused default

Core ML: pin compute_unit=CPU_ONLY, drop the fp32 variant

Add per-backend config