align torch dn code

merge deformable dino mask-dino same code

reset norm attr

fix dino amp training

fix bbox_pred in detr postprocess

* refine swin configs and codes

* fix swin ppyoloe

* fix swin for ema and distill training

* fix configs for CI

* fix docs, test=document_fix

* add DINO

add msdeformable attention cuda op

fix export bug

* fix decoder norm

* Add DETR

* drop return_pad_mask in PadBatch