[Feature] Add passes that add mkldnn related attributes instead of changing native ops

[lidanqing-intel]

Use passes to add mkldnn related attributes instead of changing native ops
@tsocha

[paddle-bot-old]

您好，我们已经收到了您的问题，会安排技术人员尽快解答您的问题，请耐心等待。请您再次检查是否提供了清晰的问题描述、复现代码、环境&版本、报错信息等。同时，您也可以通过查看官网API文档、常见问题、历史Issue、AI社区来寻求解答。祝您生活愉快～

Hi! We've received your issue and please be patient to get responded. We will arrange technicians to answer your questions as soon as possible. Please make sure that you have posted enough message to demo your request. You may also check out the API，FAQ，Github Issue and AI community to get the answer.Have a nice day!

[jakpiase]

Is there any example on how can we add attributes in inference stage?

[lidanqing-intel]

@tsocha is working on the design you can discuss with him. We need to get more info from Baidu about compatibility requirements.

[lidanqing-intel]

Meeting notes

1. Not removing any attributes that already exist in native op
2. Serilizartion: If you follow register_op steps, it will be serilizaed
3. About naming conv_oneDNN_op.cc

Question1: GPU training. CPU mkldnn training is working now. If adding this new mkldnn custom op, will it work on GPU training.This custom op is through passes, and training doesn't have such passes, so it should not affect. Is it callable from python. If it is callable it will influence training. Tomasz answer is no such plan for this.

Question2: How does this mkldnn custom op support eager mode which is most popular.

[lidanqing-intel]

related to #36541