Created by: lidanqing-intel

With this patch we do the following:

• separate QAT1.0 and QAT2.0 code - the QAT documentation update will come in a separate PR,
• make QAT2.0 use plain FP32 model to compare its accuracy with an INT8 model derived from the QAT model,
• keep the --quantized_ops option in the QAT2.0 tests - a detailed explanation on how to use it will come in the QAT documentation update.

This is Part 1 of the code changes. Part 2 comes with the PR https://github.com/PaddlePaddle/Paddle/pull/23284.