Add Resnet50 for bf16 mixed precision

[AshburnLee]

PR功能

PR支持了GPU单卡 Resnet50 的bfloat16混合精度训练。

方法

1. 添加参数use_amp_bf16 & use_pure_bf16 并设false为初始值。
2. 在train.py中调用接口decorate_bf16()，并指定使用bfloat16精度的op，将其放入custom_bf16_list中。
3. 添加执行bf16 混合精度训练的脚本Resnet50_bf16.sh

[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.}