Some of the tests are being killed because we don't produce any output
when running parallel tests in a layer with -v1.

The amount of warnings made it hard to see real failures.

Also use more recent PyPy, and test on Py 3.6 (which produces the
'async' warnings we'll use to fix #104)