Due to a change [1] after beta1, this image has altered somewhat.

[1] matplotlib/matplotlib#19619

The test tolerance is too high, and random other changes have crept in
so that some show numbers that are completely different from the
reference image.

* 726ebac changed the `Mercator`
  projection in `test_set_xyticks` to use the default globe, which
  changed the text of the ticks from degree-ish to metres.
* Ticks are quite clearly requested at 30° increments in
  `test_set_yticks_cylindrical`, so they definitely should not be at
  ±31/75°.
* Also, fix some small inconsistencies in the tests.

Fixes SciTools#1880

added missing shapefile group to download script

Due to a change [1] after beta1, this image has altered somewhat.

[1] matplotlib/matplotlib#19619

instead of hard-coding PlateCarree.

FIX: pcolor shading with nearest

Backport PRs for 0.20.1

Address Shapely 2.0 Deprecations

This adds a cache mechanism to creating Transformer objects, which
can be slow when called frequently. Additionally, this requires
adding attributes to the Mercator creation as the x/y limits are
created in the initializer which uses the cache/hash function.

…-v0.20

Cache transformer crs for v0.20.x

This was missed in the pcolormesh refactor causing some incorrect
bounds calculations when plotting the data.

Natural Earth version 5 updated some features, which caused
this test to fail the tolerance. Update the image with the new
features and remove the old forced edgecolor.

There is some conda-forge issue with Shapely/geos
so switch over to Linux for the min dependency tests.

This branch has a different set of updates, causing some of the
test tolerances to be changed. Two of these natural earth tests
are different because the shapefiles have recently been updated
and some features are different now. (For instance, the lakes
in the shapefile are now different)

The negotiation with a tile provider needed more up-to-date
TLS/SSL protocols, so for now we just pin to a lower openssl
on the runners.

Partial cherry-pick of 579be04
to increase the test tolerance of the images for MPL3.5 tests.

Backport recent master commits to v0.20.x

Remove use of setuptools_scm_git_archive

Change to assert_allclose instead of assert_almost_equal to test
small changes as recommended by Numpy's testing docs. This is due
to small tolerance changes with PROJ 9.0.1.

These only minimally changed with a shifting of the dashes in the
gridlines, but no substantive changes.

Depending on the values input to Nightshade, the floating point
precision could put the range for arccos outside of [-1, 1], which
in turn numpy returns as nan's, yielding bad geometries further
downstream. This patch clips the arccos calculations to [-1, 1] to
guarantee we aren't out of the valid floating point bounds.

Apparently the images change between PROJ 9.0.0 and 9.0.1

Backports for 0.20.x