Googlenet support bn to fit LR linear scale rule with large batch_size

[NascentAscension]

关于实验https://github.com/PaddlePaddle/PaddleClas/issues/2731的结论，添加BN层可以防止线性放大BS与LR时大幅掉精度。
可以通过启动指令输入
--override Arch.insert_bn=True(False)
控制googlenet是否插入BN层

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}