Object Detection: Add support for pretrained `weights`

[adamjstewart]

Summary

We should add support for using pretrained backbones to our semantic segmentation and object detection trainers.

Rationale

Our classification/regression/BYOL trainers support pretrained weights, we should be uniform.

Implementation

Need to figure out if any of these models support the use of timm for backbones or whether or not we can manually replace the backbone.

Alternatives

We currently support ImageNet pre-trained backbones for a limited subset, but that isn't the same.

Additional information

No response

[adamjstewart]

Update: segmentation weights are now supported but not detection weights.