Created by: wojtuss

With this patch, when running the Ernie QAT INT8 comparison test, a user can use the --fp32_model option to provide an FP32 model for accuracy measurement and comparison with an INT8 model obtained from the QAT model (passed by --qat_model option). If the --fp32_model option is not used (the default in QAT INT8 tests), the QAT model accuracy will be measured and compared.

It also adds printing average accuracy result after each (FP32 and INT8) inference run.

test=develop