Add support of the `test` split on ImageNet

[datumbox]

🚀 The feature

The new Dataset API allows returning meta-data about the images. This means that the datasets can also be used for producing predictions. As a result now it makes sense for the ImageNet dataset to also support the test split which doesn't contain labels but instead is a flat list of files.

Motivation, pitch

See this discussion for more info: #4845 (comment)

cc @pmeier @bjuncek

[pmeier]

Unfortunately, I don't have access to the test data, so I need more information to add this.

• Train and validation split come in archives:
  

  vision/torchvision/prototype/datasets/_builtin/imagenet.py

  Lines 57 to 66 in b47ea83

  	if config.split == "train":
  	images = HttpResource(
  	"ILSVRC2012_img_train.tar",
  	sha256="b08200a27a8e34218a0e58fde36b0fe8f73bc377f4acea2d91602057c3ca45bb",
  	)
  	else: # config.split == "val"
  	images = HttpResource(
  	"ILSVRC2012_img_val.tar",
  	sha256="c7e06a6c0baccf06d8dbeb6577d71efff84673a5dbdd50633ab44f8ea0456ae0",
  	)

  
  Following the pattern, is the test archive ILVSRC2012_img_test.tar?
• If yes, is it modeled like the validation archive? If yes, the images should be directly in the archive and probably named ILSVRC2012_test_00000001.JPEG.

[datumbox]

Yes there is a ILVSRC2012_img_test.tar. See these old instructions to get a better understanding on its structure.

You can get access to the archieve by signing up for a research access and then download the data from here.

Note that the test tar doesn't have labels and indeed the files are named like ILSVRC2012_test_00000001.JPEG.