[prototype] Speed up adjust_sharpness_image_tensor

[datumbox]

Related to #6818

Performance optimization for the adjust sharpness kernel:

[----------- adjust_sharpness_image_tensor cpu torch.float32 -----------]
                         |  adjust_sharpness_image_tensor old  |  fn2 new
1 threads: --------------------------------------------------------------
      (16, 3, 400, 400)  |                 274                 |    230  
      (3, 400, 400)      |                   4                 |      4  
6 threads: --------------------------------------------------------------
      (16, 3, 400, 400)  |                 300                 |    260  
      (3, 400, 400)      |                   5                 |      5  

Times are in milliseconds (ms).

[----------- adjust_sharpness_image_tensor cuda torch.float32 ----------]
                         |  adjust_sharpness_image_tensor old  |  fn2 new
1 threads: --------------------------------------------------------------
      (16, 3, 400, 400)  |                 440                 |    382  
      (3, 400, 400)      |                 150                 |    100  
6 threads: --------------------------------------------------------------
      (16, 3, 400, 400)  |                 400                 |    400  
      (3, 400, 400)      |                 150                 |    100  

Times are in microseconds (us).

[------------ adjust_sharpness_image_tensor cpu torch.uint8 ------------]
                         |  adjust_sharpness_image_tensor old  |  fn2 new
1 threads: --------------------------------------------------------------
      (16, 3, 400, 400)  |                 280                 |    240  
      (3, 400, 400)      |                   5                 |      4  
6 threads: --------------------------------------------------------------
      (16, 3, 400, 400)  |                 300                 |    260  
      (3, 400, 400)      |                   7                 |      6  

Times are in milliseconds (ms).

[------------ adjust_sharpness_image_tensor cuda torch.uint8 -----------]
                         |  adjust_sharpness_image_tensor old  |  fn2 new
1 threads: --------------------------------------------------------------
      (16, 3, 400, 400)  |                 520                 |    459  
      (3, 400, 400)      |                 190                 |    110  
6 threads: --------------------------------------------------------------
      (16, 3, 400, 400)  |                 500                 |    460  
      (3, 400, 400)      |                 180                 |    110  

Times are in microseconds (us).

cc @vfdev-5 @bjuncek @pmeier

[pmeier]

One question otherwise LGTM if CI is green.

[pmeier]

To ease review, here is the old implementation:

vision/torchvision/transforms/functional_tensor.py

Line 803 in bf58902

def _blurred_degenerate_image(img: Tensor) -> Tensor:

[pmeier]

Can push a change like this to _blend or is this a special case here?

[datumbox]

It's a special case :( We can do this only because we are allowed to subtract image1 from image2 in place. In all other cases where _blend() is used, we rely on broadcasting do that's not possible.