Speedup InceptionV3 and GoogleNet tests on windows (upgrade scipy)

[datumbox]

Rough improvement: ~96%

• Before: 931.24s - reference

327.29s call     test/test_models.py::ModelTester::test_inception_v3_cpu
326.23s call     test/test_models.py::ModelTester::test_inception_v3_cuda
135.34s call     test/test_models.py::ModelTester::test_googlenet_cuda
133.91s call     test/test_models.py::ModelTester::test_googlenet_cpu
5.43s call     test/test_models.py::ModelTester::test_inceptionv3_eval
3.04s call     test/test_models.py::ModelTester::test_googlenet_eval

• After: 32.38s - reference

7.97s call     test/test_models.py::ModelTester::test_googlenet_cuda
7.41s call     test/test_models.py::ModelTester::test_inception_v3_cpu
5.24s call     test/test_models.py::ModelTester::test_inception_v3_cuda
4.52s call     test/test_models.py::ModelTester::test_googlenet_cpu
4.50s call     test/test_models.py::ModelTester::test_inception_v3_eval
2.74s call     test/test_models.py::ModelTester::test_googlenet_eval

The above are rough estimations based on a couple or runs.

The problem is traced to the use of scipy in the weight initialization of the two models:

vision/torchvision/models/inception.py

Lines 122 to 124 in 90645cc

	import scipy.stats as stats
	stddev = m.stddev if hasattr(m, 'stddev') else 0.1
	X = stats.truncnorm(-2, 2, scale=stddev)

vision/torchvision/models/googlenet.py

Lines 126 to 128 in 90645cc

	import scipy.stats as stats
	X = stats.truncnorm(-2, 2, scale=0.01)
	values = torch.as_tensor(X.rvs(m.weight.numel()), dtype=m.weight.dtype)

The slow down is caused by a regression on scipy affecting versions 1.4.x: scipy/scipy#11299. The reason why only Windows runs are affected is because our CI config is pinning the version of scipy to 1.4.1 on Windows but not on Linux. This was introduced at #2328 (first commit of the file) and it's probably by accident that the two files are not aligned.

By unpinning the version and allowing scipy to use newer versions on Windows, the issue is fixed. This is an alternative solution to #3195.

[codecov]

Codecov Report

Merging #3196 (f3943d5) into tests/speedup (1a300d8) will not change coverage.
The diff coverage is n/a.

@@              Coverage Diff               @@
##           tests/speedup    #3196   +/-   ##
==============================================
  Coverage          73.48%   73.48%           
==============================================
  Files                 99       99           
  Lines               9230     9230           
  Branches            1476     1476           
==============================================
  Hits                6783     6783           
  Misses              1991     1991           
  Partials             456      456           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 1a300d8...f3943d5. Read the comment docs.

[datumbox]

The failing tests are not related to this PR.

[fmassa]

Awesome, thanks for the PR @datumbox !