Avoid a 2nd forward pass during validation step

[bw4sz]

Deepforest primary connects to Torchvision models. Torchvision models have a 'train' and 'eval' mode, which are uncomfortably named, because it also effects what is outputted. In train mode, the loss is returned, in eval mode, the predictions are returned.

https://discuss.pytorch.org/t/how-to-get-losses-and-predictions-at-the-same-time/167223

from https://pytorch.org/vision/main/models/generated/torchvision.models.detection.retinanet_resnet50_fpn.html#torchvision.models.detection.retinanet_resnet50_fpn

The line of code is here.

DeepForest/src/deepforest/main.py

Line 677 in e5f3301

preds = self.model.forward(images)

1. its a performance hit to run the network twice.
2. When we start to add more architectures we don't want to enforce this structure
3. Its got a terrible name and we always need to make sure we wrap in no_grad to make sure it doesn't update weights or batch_norms.