Bit width of image

I think it would be convenient to have if enough users require it (tbh I saw very few posts demanding it) and make the torch API more consistent with other python packages.

My specific use case only cares about torch.from_numpy():

if img_np.dtype == np.uint16:                # get rid
    img_np = img_np.astype(np.int32)         # of this part 

img_t = torch.from_numpy(img_np).type(torch.FloatTensor)