[Feature] Support DiffusionDet Inference in Projects

[BIGWangYuDong]

Ready for the inference-time performance of DiffusionDet in MMDetection projects matches that in the original paper(repo)

Note:
During inference time, DiffusionDet will randomly create boxes, which may affect the results (AP).

If hard-setting seeds=0 before generating random boxes, the AP will be same:

# hard set seed=0 before generating random boxes
seed = 0
random.seed(seed)
# np.random.seed(seed)
torch.manual_seed(seed)
# torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)

...

noise_bboxes_raw = torch.randn(
    (self.num_proposals, 4),
    device=device)

...

repo	mAP
DifussionDet (without any changing)	45.5
DifussionDet (hard set seed=0)	45.66
mmdet repo (hard set seed=0)	45.7
mmdet repo (random seed)	45.6~45.8

In addition, we give a script to convert DiffusionDet-resnet50 to MMDet version.

python projects/DiffusionDet/model_converters/diffusiondet_resnet_to_mmdet.py ${PATH_DOWNLOAD_FROM_DIFFUSIONDET} ${CONVERTED_SAVING_PATH}