Also added a fix for residual issues from pytest-dev/py#169, which is necessary until either the 1.5.0 release is dropped, the trove classifiers get fixed, or whatever test dependencies are including py start pinning on Python 3.3 themselves.

OK, so while testing this, I realized that we're not being hit by that codecov issue. What was happening was that codecov run -m pytest -v was inferring that the only source was dateutil.test for whatever reason.

By explicitly adding coverage run --source=dateutil -m pytest -v, we get the full report.

Unfortunately, this doesn't actually bring up our coverage at all, because the old erroneous reports from Appveyor were counting dateutil.test twice, and dateutil.test has excellent coverage, as you would expect from a test suite...