Refactor & enable JIT tests in all models and add warnings if skipped

[datumbox]

Fixes #3028

Initial commits of this PR should have caused the following models to fail:

• googlenet
• inceptionv3
• maskrcnn

The failures were related to non JIT problems on the tests but we could not see them because the tests were not running.

This PR fixes the problems on the above tests, refactors the unit-tests and re-enables JIT tests for all models.

[datumbox]

@fmassa Unfortunately the tests did not fail which means that the JIT scripts did not run. This can be confirm by checking any of the run logs for the following warnings:

test/test_models.py::ModelTester::test_alexnet_cpu
  /Users/distiller/project/test/common_utils.py:281: RuntimeWarning: The checkModule test for AlexNet was skipped. This test checks if the module's results in TorchScript match eager and that it can be exported. To run these tests make sure you set the environment variable PYTORCH_TEST_WITH_SLOW=1 and that the test is not manually skipped.
    warnings.warn(msg, RuntimeWarning)

Any ideas on how we can turn on the PYTORCH_TEST_WITH_SLOW env var at CI? Alternatively we can hardcode it here until this is fixed:

vision/test/common_utils.py

Lines 34 to 35 in b2c0b3d

	TEST_WITH_SLOW = os.getenv('PYTORCH_TEST_WITH_SLOW', '0') == '1'
	# TEST_WITH_SLOW = True # TODO: Delete this line once there is a PYTORCH_TEST_WITH_SLOW aware CI job

[fmassa]

Thanks a lot for looking into this, we should get this fixed quickly.

I think we should try to switch the flag to always run the slow tests, and we should see if CI timeout or not.
If you want to set the env var for tests in CI only, you can add it in

vision/.circleci/unittest/linux/scripts/run_test.sh

Line 8 in a884cb7

python -m torch.utils.collect_env

and

vision/.circleci/unittest/windows/scripts/run_test.sh

Line 8 in a884cb7

python -m torch.utils.collect_env

It's likely to fail for models:
maskrcnn

Just note that we have a separate test checking detection models that they run on torchscript, so we should be safe there, see

vision/test/test_models.py

Lines 213 to 215 in a884cb7

	scripted_model = torch.jit.script(model)
	scripted_model.eval()
	scripted_out = scripted_model(model_input)[1]

[fmassa]

Thanks!