NOMRG refactor test_models to use pytest

[NicolasHug]

Just a small POC

@fmassa, this is how we could factorize test_models if we relied on pytest. In fact we could factorize much more and get rid of the unittest.TestCase dependency, but this would be more work.

[NicolasHug]

This allows to preserve the use of python tests/test_models.py

Most packages don't do this and just rely on calling pytest or python -m pytest from the command line instead of python.

[codecov]

Codecov Report

Merging #3528 (fff4683) into master (414427d) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #3528   +/-   ##
=======================================
  Coverage   78.75%   78.75%           
=======================================
  Files         105      105           
  Lines        9748     9748           
  Branches     1565     1565           
=======================================
  Hits         7677     7677           
  Misses       1581     1581           
  Partials      490      490           

---

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 414427d...fff4683. Read the comment docs.

[datumbox]

@pmeier you might be interested in this.

[NicolasHug]

@pmeier you might be interested in this.

For context this is just a POC where I'm trying to showcase the benefits of using pytest in torchvision (#3505 is another one with parameterized). There's still a lot of work/discussion to do before we can start adopting pytest for good. I'm currently working on making it run internally on fb. I'll keep updated on my progress but it's still very preliminary.

[NicolasHug]

Closing in favor of #3697