FIX: plot after contour shouldn't shrink extent

[rcomer]

Rationale

Fixes #2128.

We already have the ignore=False within update_from_data_xy so that the contour call itself doesn't shrink the extent. The equivalent line for plot is

https://github.com/matplotlib/matplotlib/blob/d731d48b2431da0cf8d7edaa687f997857b112af/lib/matplotlib/axes/_base.py#L2366-L2368

which relies on ax.ignore_existing_data_limits to decide whether to shrink. Under normal circumstances, this flag is set within update_datalim, which is called by Matplotlib's contour. However, for the combination of transform and invalid latitude in the linked issue, the xys that gets passed to update_datalim is all nans.

Implications

[rcomer]

The Ubuntu py3.8 tests seem to be picking up an older version of proj 😕

[greglucas]

Thanks for the clear explanation, I think I'm following now!

Should we actually be using this Axes method instead? Which I think basically does what you're doing here.
https://github.com/matplotlib/matplotlib/blob/d772e0d4306eac3a1b14f9099711f6261153a1f0/lib/matplotlib/axes/_base.py#L2485-L2508

On the tests... :( We pin to that version of proj/pyproj, but it is odd that it is failing on only a couple of tests.

[rcomer]

I looked at the py38 CI on another recent PR, and it has proj-8, but this is getting proj-7.2.

[greglucas]

I thought proj was only being used for geodesics at this point, but perhaps that is the issue... I'm still hopeful that we'll rely exclusively on pyproj in the future so we don't have separate pins for the two.

proj8 was released in March 2021, so we could probably pin that high if need be until we get the proj dependence removed: https://github.com/OSGeo/PROJ/releases

It also looks like we need to update something with the sphinx gallery logo for the docs too. Lets address those in separate PRs though.

[greglucas]

The failling tests and docs are unrelated and have been addressed in other PRs.

[rcomer]

Thanks @greglucas!