Refactor _RandomNCrop to RandomCrop + ExtractPatches

[isaaccorley]

This PR fixes #2667 and refactors _RandomNCrop out of segmentation and change detection datamodules in favor of K.RandomCrop during training and ExtractPatches for val/test/predicting.

[adamjstewart]

I would love to get rid of _ExtractPatches too, but I'm guessing kornia/kornia#1300 prevents us from doing that? It may be time to work more closely with the Kornia developers to upstream and replace our broken custom transforms.

[isaaccorley]

I would love to get rid of _ExtractPatches too, but I'm guessing kornia/kornia#1300 prevents us from doing that? It may be time to work more closely with the Kornia developers to upstream and replace our broken custom transforms.

I'll bring that back from the grave after this gets merged. I want to see if they can help me with extracting patches but also being able to extract bboxes since this is important for object detection datasets that have large imagery. We don't want to lose users to SAHI.

[adamjstewart]

Due to the bugs for batch_size=1 and lack of support for bboxes, I'm strongly considering using RandomCrop for train and CenterCrop for val/test (with large patch size and small batch size) instead of replacing our hacky _RandomNCrop with another hacky _ExtractPatches. Then we can find time to contribute something better upstream to Kornia. How would you feel about that?

[isaaccorley]

I think _ExtractPatches is fine and likely more correct, particularly for datasets like LEVIR-CD where patch size is typically 256x256 and the image size is 1024x1024 so your validation and test metrics will actually encompass the entire test set images.

Using CenterCrop this would not be the case and not be able to provide an apples to apples comparison with other baselines.