FIX: Deepcopy of CRS with str/epsg

[lgolston]

Rationale

When using initializing crs.CRS with a string, or calling crs.epsg, there are errors when loading from a pickled copy or when performing deepcopy. This PR updates reduce to get those processes working.

Implications

Closes #2042 and the new issue recently reported in #1336

Checklist

Two previously failing cases are added to the crs test fixture.

[lgolston]

The minimum packages run failed, and now I see pyproj has a related bug that was fixed in version 3.2.0 (pyproj4/pyproj#901). Would bumping the minimum version from 3.1.0 to 3.2.0 be okay?

[greglucas]

Yes, I think that should be fine, we've been trying to do roughly 2 years for dependencies and that was released in 2021.
https://github.com/pyproj4/pyproj/releases/tag/3.2.0

[lgolston]

Okay, I test locally and v3.3.1 is actually minimum needed for this, and have pushed a new update.
https://github.com/pyproj4/pyproj/releases/tag/3.3.1

[dopplershift]

I'm happy to go with Proj 3.3.1 since that's at 23 months old.

[lgolston]

I realized that checking if self.proj4_params == {} was a little myopic. That made deepcopy work if the proj4_params input to crs.CRS was a string, but still did not work for a dict. The PR is now updated to handle both cases.

pyproj is bumped as before (and one place matplotlib to 3.5 was missed in #2276)

As an aside, I also think getstate only needs to pop the _local variable, since it is unpickable, but just left it alone since it works either way.

There seem to be intermittent / unrelated failures recently that go away upon re-running (here with windows-latest, 3.9, false), which I do not understand.

[lgolston]

Re-ran a fourth time and all tests are passing this time.
Run #1 there were a lot of errors in windows 3.9, I saw something similar when my conda environment was a little messed up before.
Run #2 and #3 there was a HTTP 502 error in downloading a few tiles from Google.
I think it is ready to go, but let me know if there are any issues.