Extend RandomShortestSize to support Video specific flavour of the augmentation

[datumbox]

Our RandomShortestSize implementation on the references is developed with Object Detection in mind. Videos require a slight variation (see reference implementation for detection vs video).

This PR extends the transform in a BC-way so that it can support both. In particular, we make the max_size optional and this allows us to reparameterize the transform for videos as such:

from torchvision.prototype.transforms import RandomShortestSize
import math
import torch

x = torch.randn(7, 11, 3, 450, 800)
t = RandomShortestSize(list(range(256, 320+1)))
z = t(x)
print(z.shape)

size = min(z.shape[-2:])

_, t, c, h, w = x.shape
if w < h:
    new_h = int(math.floor((float(h) / w) * size))
    new_w = size
else:
    new_h = size
    new_w = int(math.floor((float(w) / h) * size))

print(new_h, new_w)

assert (new_h, new_w) == tuple(z.shape[-2:])

Though the names of min_size and max_size are confusing (better names would have been shortside_min_size_range and longside_max_size), their semantics align with the arguments that F.resize() has for size and and max_size. On the latter the default value of max_size (which again applies to the longest edge) is None, so this transform uses the same semantics and default values as in other places of the API.

[vfdev-5]

OK to me, thanks @datumbox !