Consider reverting optimization in normalization

[jaraco]

In #533 (comment)_, I got the impression that the optimization may be premature. Now that code bases have stabilized and the performance checks are more precise, let's revisit 3e7f3ac and decide if it's really worth the complication.

[jaraco]

As a baseline, here are a couple of snapshots of the performance tests without any diff:

exercises.py:cached distribution: 128 µsec (-2 µsec, -2%)
exercises.py:discovery: 129 µsec (-1 µsec, -1%)
exercises.py:entry_points(): 1.47 msec (+40 µsec, 3%)
exercises.py:entrypoint_regexp_perf: 48.7 µsec (0 nsec, 0%)
exercises.py:normalize_perf: 93.3 nsec (+2.1 nsec, 2%)
exercises.py:uncached distribution: 215 µsec (-5 µsec, -2%)

exercises.py:cached distribution: 131 µsec (0 nsec, 0%)
exercises.py:discovery: 132 µsec (-2 µsec, -1%)
exercises.py:entry_points(): 1.45 msec (-10 µsec, -1%)
exercises.py:entrypoint_regexp_perf: 49.5 µsec (+200 nsec, 0%)
exercises.py:normalize_perf: 87.8 nsec (-1.3 nsec, -1%)
exercises.py:uncached distribution: 220 µsec (+1 µsec, 0%)

I've seen even higher numbers in the 3-4% range.

So there's quite a bit of jitter in the tests, meaning we'd probably want to see 6% or better improvement to justify a change.