Move from PIL to torchvision.io.decode_image

[ebsmothers]

Instead of using PIL.open, @NicolasHug has pointed out that we can move to torchvision.io.decode_image to speed up image processing by doing everything on pure tensors (see also. This would also allow us to drop our explicit PIL requirement). This should entail:

1. changing load_image to use torchvision.io.decode_image
2. updating CLIPImageTransform to accept tensors instead of PIL.Image (in the short term we can keep PIL.Image support for backwards compatibility)

[Ankur-singh]

@ebsmothers Can you please assign this to me?

Thank you 😁

[ebsmothers]

@Ankur-singh just assigned it to you, thanks for taking it on!

[NicolasHug]

Feel free to ping me for help / reviews / best practices!