Quantizing models for edge deployment - INT8 vs INT4 vs mixed precision?