os.environ (setting envar COLUMNS), which at least caused
test_float_default() to fail if the tests were run more than once.

This repairs the test_optparse -R failures Neal reported on
python-dev.  It also explains some seemingly bizarre test_optparse
failures we saw a couple weeks ago on the buildbots, when
test_optparse failed due to test_file failing to clean up after
itself, and then test_optparse failed in an entirely different
way when regrtest's -w option ran test_optparse a second time.
It's now obvious that make_parser() permanently changing os.environ
was responsible for the second half of that.