tests: cleanup and ci hardening

[henryiii]

Refactor the tests. Use a consistent failing/skipping scheme.

• Attempt to allow 3.9-dev to fail (not sure how it works)
• Fail-fast diagnostic removed, fails fast (once) with the correct error message without a custom wrapper.
• Using xfail where possible (skip means not-applicable, xfail means it should ideally pass)
• Use importorskip instead of ugly custom skips
• PyTest 3.1+ required instead of 3.0 (current is 6.0 for Py3, 4.6.9 for Py2) - better warning reporting
• Use info.py to simplify skips/xfails/checks
• Update links to PyPy issues
• Several (7 or so) skipped PyPy tests are now passing (especially those requiring NumPy) (reported by xfail!)
• Strings used for the skipif/xfails in most cases so that PyTest reports the condition causing the skip/xfail directly in the report.

Skips/xfails are more consistent, no more pytest.bug_in_pypy, pytest.unsupported_on_pypy_lt_6, etc. (PyPy 7+ required, so ignoring check for old PyPy verisions). No more injection into the pytest module. xfails are reported at the end and most are strictly required. We can use reasons or raises to improve the test reporting.

[bstaletic]

I really like this PR.

[henryiii]

I believe 7 or so tests are now passing in PyPy that used to fail (and were hidden by skip instead of xfail), mostly due to better NumPy support in recent PyPy's.

[YannickJadoul]

I believe 7 or so tests are now passing in PyPy that used to fail (and were hidden by skip instead of xfail), mostly due to better NumPy support in recent PyPy's.

That's great to hear! :-)

[henryiii]

If I get two approvals, I'll squash and merge. :)

[bstaletic]

This looks great. Thanks for the PR.

[rwgk]

Awesome, thanks Henry!