Add int8 encoder variant (~half size, accuracy-neutral)

Add SenseVoiceSmall CoreML (preprocessor fp32 + encoder fp16/ANE + fp32 fallback + vocab + card)

initial commit