Update backbones to use default PyTorch implementations

[busycalibrating]

Currently, the primary ResNet backbone uses a custom implementation that would not be directly compatible with the default PyTorch implementations (in addition to having a slightly different layer configuration). I think it would be advantageous to move to using the standard models offered in torchvision as most people likely default to those.

[MalteEbner]

I agree that using the standard models offered in torchvision would be better than having our own models. What we are currently thinking about is a good migration strategy, as the current checkpoints are only compatible with the current resnet models, but not necessarily with the default pytorch ones.

[IgorSusmelj]

Would keep the old ResNet structure for datasets such as cifar10. Images of size 32x32 don't work well with the standard ResNets. The 7x7 conv with stride=2 and maxpool reduce the feature a lot.

[guarin]

We switched to the default torchvision resnets in all our model examples.