Detection models weights

Pretrained weights for detection models.

EfficientDet models were trained by Google, then ported to PyTorch by @zylo117 in his repository and then mapped to models in this repo by @bonlime . Mapping gives absolute error < 1e-8 for raw outputs.
upd. additionally removed extra bias in first BiFPN layer downsample convs. Pretrained weights for them were ~1e-5. So it doesn't affect validation results.

RetinaNet models are ported from mmdetection. Mmdetection ResNet is slightly different (stride 2 in conv1x1 instead of conv3x3) and order of anchors is also different so it's impossible to do inference using this weights but they work much better for transfer learning than starting from imagenet pretrain