[proto] Small optimization for gaussian_blur functional op

[vfdev-5]

Description:

• use softmax and multiplication to speed-up gaussian blur op

[-------------- Gaussian Blur cpu torch.uint8 --------------]
                     |  gaussian_blur v1  |  gaussian_blur v2
1 threads: --------------------------------------------------
      (3, 400, 400)  |        2.78        |        2.65
6 threads: --------------------------------------------------
      (3, 400, 400)  |        1.08        |        1.01

Times are in milliseconds (ms).

[-------------- Gaussian Blur cuda torch.uint8 -------------]
                     |  gaussian_blur v1  |  gaussian_blur v2
1 threads: --------------------------------------------------
      (3, 400, 400)  |        342         |        227
6 threads: --------------------------------------------------
      (3, 400, 400)  |        259         |        227

Times are in microseconds (us).

[------------- Gaussian Blur cpu torch.float32 -------------]
                     |  gaussian_blur v1  |  gaussian_blur v2
1 threads: --------------------------------------------------
      (3, 400, 400)  |        2160        |        2080
6 threads: --------------------------------------------------
      (3, 400, 400)  |         917        |         840

Times are in microseconds (us).

[------------- Gaussian Blur cuda torch.float32 ------------]
                     |  gaussian_blur v1  |  gaussian_blur v2
1 threads: --------------------------------------------------
      (3, 400, 400)  |        232         |        200
6 threads: --------------------------------------------------
      (3, 400, 400)  |        231         |        200

Times are in microseconds (us).

Code source

---

This PR may also help with flaky gaussian blur tests.

Context:

Gaussian blur eager vs jit tests are flaky for uint8 input. This may be due to the fact that we cast to float32, perform the opertation in float32 and cast back to uint8. We generate conv kernels (created using exp op) in float32 and maybe accumulating precision errors.

Refs: #6755

[vfdev-5]

Successful runs on ubuntu-latest:

• https://github.com/pytorch/vision/actions/runs/3263884056/attempts/1
• https://github.com/pytorch/vision/actions/runs/3263884056/attempts/2
• https://github.com/pytorch/vision/actions/runs/3263884056/attempts/3
• https://github.com/pytorch/vision/actions/runs/3263884056

---

(Before optimization trick)

• https://github.com/pytorch/vision/actions/runs/3241542307/attempts/1
• https://github.com/pytorch/vision/actions/runs/3241542307/attempts/2
• https://github.com/pytorch/vision/actions/runs/3241542307

[pmeier]

LGTM if CI is green!

[github-actions]

Hey @vfdev-5!

You merged this PR, but no labels were added. The list of valid labels is available at https://github.com/pytorch/vision/blob/main/.github/process_commit.py