Defer import of zipp

[danielhollas]

Impoting zipp takes between 2-4 ms on my machine, which is at least 10% of overall import time of the importlib_metadata package (depending on Python version). Since this package is only used in a single code path it makes sense to defer it.

To estimate the cost of repeatedly calling import zipp I ran

❯ python -m timeit -s "import zipp"  "import zipp"
5000000 loops, best of 5: 59.9 nsec per loop

(this is with cpython main branch)

Since this is in a codepath that reads a zipfile from disk, this overhead is negligible.

[jaraco]

I'm starting to feel like a frog being slowly boiled. I'm seeing repeated optimizations like this being handled piecemeal, and each time I do, I need to go through the same process: point out that there's a systemic problem that has no solution, ensure there's a comment protecting the change and linking to the motivation, then merge and roll out the change. </rant>

Looks good. I'll add the comment linking back to the motivation and analysis (here). Thanks for the contrib.