Update Image training code for downstream datasets (iWildcam, transfer datasets)

[ohaijen]

Update trochvision/train.py to accomodate:

• iWildcam. This code would be read through the ImageFolder format, but requires different resizes and crops.
• Transfer datasets (Aircraft, DTD, Flowers) - these would be read through the TorchVision API and cached in /tmp. These don't need any new transforms, just the class loading.