Add ConViT with pretrained weights

[amaarora]

No description provided.

[amaarora]

@rwightman

Validation Top-1 accuracies (match with Table 1 in paper):

ConViT Tiny: 73.116
ConViT Small: 81.418
ConViT Base: 82.288

NOTE: ConViT Base is 0.12% less than the 82.4% reported in the paper.

All validation numbers are for 224x224 image size.

[rwightman]

@amaarora do you think it's ready for me to pull, verify, merge?

[amaarora]

@rwightman yes :)