[prototype] Gaussian Blur clean up

[datumbox]

Related to #6818

This PR:

• Cleans up the assertions on the gaussian_blur kernel
• Simplifies the reshaping logic
• Adds in-place ops where possible

No regression on the speed, just a small 5% improvement on CUDA:

[------------- gaussian_blur cpu torch.float32 -------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |    4 (+-  0) ms  |    4 (+-  0) ms
      (16, 3, 400, 500)  |  306 (+-  3) ms  |  306 (+-  2) ms
6 threads: --------------------------------------------------
      (3, 400, 500)      |    6 (+-  0) ms  |    6 (+-  0) ms
      (16, 3, 400, 500)  |  334 (+-  1) ms  |  334 (+-  4) ms

Times are in milliseconds (ms).

[------------- gaussian_blur cuda torch.float32 ------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |  119 (+-  1) us  |  112 (+-  1) us
      (16, 3, 400, 500)  |  266 (+-  0) us  |  266 (+-  0) us
6 threads: --------------------------------------------------
      (3, 400, 500)      |  119 (+-  2) us  |  113 (+-  2) us
      (16, 3, 400, 500)  |  266 (+-  1) us  |  266 (+-  0) us

Times are in microseconds (us).

[-------------- gaussian_blur cpu torch.uint8 --------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |    5 (+-  0) ms  |    5 (+-  0) ms
      (16, 3, 400, 500)  |  355 (+-  1) ms  |  331 (+-  1) ms
6 threads: --------------------------------------------------
      (3, 400, 500)      |    7 (+-  0) ms  |    7 (+-  0) ms
      (16, 3, 400, 500)  |  383 (+-  2) ms  |  359 (+-  5) ms

Times are in milliseconds (ms).

[-------------- gaussian_blur cuda torch.uint8 -------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |  150 (+-  1) us  |  142 (+-  1) us
      (16, 3, 400, 500)  |  423 (+-  0) us  |  423 (+-  0) us
6 threads: --------------------------------------------------
      (3, 400, 500)      |  150 (+-  3) us  |  142 (+-  3) us
      (16, 3, 400, 500)  |  423 (+-  0) us  |  423 (+-  0) us

Times are in microseconds (us).

cc @vfdev-5 @bjuncek @pmeier

[vfdev-5]

OK to me. I'm not sure why we can't use _cast_squeeze_in, _cast_squeeze_out anymore, but ok.

[datumbox]

@vfdev-5 We could use it. But this means we will have multiple pieces of code handling reshaping (or needs_unsquash and need_squeeze in the previous code). In addition the casting mechanism in _cast_squeeze_in and _cast_squeeze_out makes assumptions over the order of preference of the provided dtypes and requires unnecessary complex checks for rounding. Casting can be simplified; the only thing we need to check is if the input was float and then just round and cast.

[datumbox]

It seems that this closed the gap between V1 and V2 for the GaussianBlur transform. My new benchmarks between V1+pure tensor and V2+feature report:

[------------- gaussian_blur cpu torch.float32 -------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |   13 (+-  1) ms  |    4 (+-  0) ms
      (16, 3, 400, 500)  |  306 (+-  1) ms  |  306 (+-  0) ms

Times are in milliseconds (ms).

[------------- gaussian_blur cuda torch.float32 ------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |  247 (+- 37) us  |  156 (+-  1) us
      (16, 3, 400, 500)  |  356 (+-  2) us  |  269 (+-  0) us

Times are in microseconds (us).

[-------------- gaussian_blur cpu torch.uint8 --------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |    5 (+-  0) ms  |    5 (+-  0) ms
      (16, 3, 400, 500)  |  351 (+-  1) ms  |  330 (+-  1) ms

Times are in milliseconds (ms).

[-------------- gaussian_blur cuda torch.uint8 -------------]
                         |        old       |        new     
1 threads: --------------------------------------------------
      (3, 400, 500)      |  261 (+-  5) us  |  189 (+-  1) us
      (16, 3, 400, 500)  |  515 (+-  4) us  |  426 (+-  0) us

Times are in microseconds (us).

@vfdev-5 Might be worth rerunning the benchmarks later on your side to confirm.