Improve DINO

[hhaAndroid]

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily get feedback. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

Backbone	Model	Lr schd	Better-Hyper	box AP	Config	Download
R-50	DINO-4scale	12e	False	49.0	config	model | log
R-50	DINO-4scale	12e	True	50.1	config	model | log

[SyedShaQutub]

I recently discovered a significant modification in the DINO-DETR code within the mmdetection repository. Specifically, I noticed that the self-attention section of both the encoder and decoder transformers only includes normalization in the code, without the addition of identity after the multi-head attention block. Is there a particular reason for this substantial change in the code?