Make get_image_size and get_image_num_channels public

[datumbox]

The F._get_image_size() and F._get_image_num_channels() are quite useful methods to determine the size and channels of images and they are used across many classes. This PR makes them public.

cc @vfdev-5 @datumbox

[fmassa]

Changes LGTM.

Only thing I would consider if we are making it public is about the order of the elements being returned. I've always found it confusing with PIL to have the sizes be (w, h) instead of (h, w), and _get_image_size returns (w, h) to minimize the number of lines changed when we introduced Tensor support for the transforms.

I'm not sure if it's worth changing the order of returned elements, but it's something worth keeping in mind.

Thoughts?

[datumbox]

For transparency, I had an offline chat with @fmassa to see if it's worth adding a new method that returns all 3 arguments as (c, h, w) and deprecate the old. Unfortunately there are a few places we use img_size=(w,h) in public APIs:

vision/torchvision/transforms/transforms.py

Lines 1446 to 1448 in 7947fc8

	img_size = F._get_image_size(img)
	ret = self.get_params(self.degrees, self.translate, self.scale, self.shear, img_size)

Thus we will merge this version.

[praveern]

Cannot seem to be able to access these functions anymore. Error: #4328