Further optimize distribution() searching.

[anntzer]

Locally, this speeds up distribution("ipython") on a large-ish
environment by nearly 2-fold. (The GHA benchmark goes from ~1.6ms
to ~0.8ms.)

All the following changes matter for the performance gain:

• avoid needlessly reinstanciating Prepareds.
• switch from os.path.splitext (very slow) to str.rpartition
• avoid having to add back a dot to the ext part given by rpartition
• precompute egg_prefix and versionless_egg_name

(See https://gitlab.com/python-devs/importlib_metadata/-/issues/95 for
rationale why performance matters.)

[jaraco]

Thanks for this. It's excellent. The only real wish I have (for future consideration) is if these optimizations had been committed and pushed separately, we could see the impact of each change independently.

Any idea why the Python 3.6 tests are timing out?

[jaraco]

Any idea why the Python 3.6 tests are timing out?

Oh, dear. I see the Python 3.6 tests have been timing out for some time, although apparently only intermittently. I'll bet it's some bug with pips resolver and selective dependencies on self (importlib_metadata). I'll deal with that in another issue/pr.

[anntzer]

if these optimizations had been committed and pushed separately, we could see the impact of each change independently.

Done.

[jaraco]

Looks like the second and third commits had the most impact:

importlib_metadata HEAD $ git checkout -q c1af1d6~1; tox -q -e perf
200 loops, best of 5: 1.47 msec per loop
________________________________________________________________ summary ________________________________________________________________
  perf: commands succeeded
  congratulations :)

importlib_metadata HEAD $ git checkout -q c1af1d6; tox -q -e perf
200 loops, best of 5: 1.45 msec per loop
________________________________________________________________ summary ________________________________________________________________
  perf: commands succeeded
  congratulations :)

importlib_metadata HEAD $ git checkout -q c9c0909; tox -q -e perf
200 loops, best of 5: 1.14 msec per loop
________________________________________________________________ summary ________________________________________________________________
  perf: commands succeeded
  congratulations :)

importlib_metadata HEAD $ git checkout -q 7fc4f11; tox -q -e perf
200 loops, best of 5: 928 usec per loop
________________________________________________________________ summary ________________________________________________________________
  perf: commands succeeded
  congratulations :)