Various tests fail when run under coverage. A primary culprit is refcount tests
which fail as the counts are thrown off by the coverage code. A new decorator
-- test.support.refcount_test -- is used to decorate tests which test refcounts
and to skip them when running under coverage. Other tests simply fail because
of changes in the system (e.g., __local__ suddenly appearing).

Thanks to Kristian Vlaardingerbroek for helping to diagnose the test failures.