The old way, we made and deleted .pth file around each test.  This
caused problems because pth files are written to a common location
shared by all test workers, so tests would fail because pth files were
being deleted as other workers were trying to use them.

Now we make the pth file once per session.  There's still a chance that
a worker will be using a pth file just as another worker is ending and
removing it.

Eight fewer sub-processes

The old code expected a dict with int keys, but iterables are simpler, so get
rid of a lot of dict.fromkeys() nonsense.

It didn't work because the os module is frozen on 3.11, so it gets no
measurement.  Use -X to disable the frozen modules feature for the test.

The fix for CTracer is egregious and will need to be updated when there's a
supported way to do it.

The fullcoverage skip is noted in
https://github.com/nedbat/coveragepy/issues/1278

The raise_through_with skip is noted in
https://github.com/nedbat/coveragepy/issues/1270

Originally I used setup_test because something went wrong when I used setUp.  I
wrote https://github.com/pytest-dev/pytest/issues/8424 about it.  There they say
to use `-p no:nose` to disable nose interpretation.

But now I simply went back to setUp, and all seems well? Not sure what changed,
but using the expected name is better.

Python 3.11 made a traceback look like this:

    Traceback (most recent call last):
      File "{path}", line 8, in <module>
        print(sys.argv[1])
              ~~~~~~~~^^^
    IndexError: list index out of range

We needed to not care if that tilde-caret line was present or not.