I am attaching a new patch. Marcus can you apply this and verify that using the
feature flag as documented enables the same speedup you reported with your tests.

In my testing I used test_liveserver.py and ran it 6 times in a row.
I recorded the time pythonX -m pytest for the last 5 runs.

I ran the loop 4 times under:

  python2 with feature_flags={}
  python3 with feature_flags={}
  python2 with with feature_flags={"cache_tracker": ""}
  python3 with with feature_flags={"cache_tracker": ""}

The cached tracker does seem to be faster. With python2 test_liveserver.py completes
about 4% faster on average. For python3 it completes about 16% faster. I'm not seeing
your x3 improvement but that may be because the time to run the test is dominated by
pytest and setup overhead. Also I am running anydbm as the backend that may add to the
overhead.

If this looks ok, I'll get it committed and then work on how to test.

I think I might be able to set up an i/o buffer for the wsgi server to write to.
Passing the i/o buffer in via a feature flag could work. Then I should be able to
get the output just from the wsgi handler from there in the test harness.

-- rouilj