My quantized tflite model produces different results when the inference is made using the NPU+NNAPI instead of the CPU. What could be the reason for that?