Add points to geostationary bounding box when determining subset slices

[djhoese]

This is an extension of #729 that applies the same idea to general get_area_slices. I had been working with Claude on reviewing #729 and #700 as they deal with similar issues. Claude pointed out flaws in #700 and suggested that until spherical intersection calculations in pyresample are better, #729's solution would be a clean replacement for what #700 is doing.

Claude wrote this code. I asked it to make a better suite of tests that borrowed from the assertions in #700, but with more cases. One of the flaws of #700 is that it really only makes the calculations better where the target area is completely inside the source area. I wanted more cases.

I don't think #700 should be merged and instead #729 and this PR should be merged. I'll have Claude update this PR with a shared helper function once I merge #729.

• Closes #xxxx
• Tests added
• Tests passed
• Fully documented

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.98%. Comparing base (4a46786) to head (4662c4b).
⚠️ Report is 15 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #757      +/-   ##
==========================================
+ Coverage   93.90%   93.98%   +0.08%     
==========================================
  Files          89       89              
  Lines       13924    13989      +65     
==========================================
+ Hits        13075    13148      +73     
+ Misses        849      841       -8     

Flag	Coverage Δ
unittests	93.98% <100.00%> (+0.08%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[djhoese]

Performance numbers from Claude:

Overall the PR is about 1.5× slower than main across these cases, and full-disk sources are unchanged. Partial-disk sources are 1.1–3.3× slower, but in 7 of the 9 partial-disk cases main's faster answer was a slice that was too small (or an error).

The table compares main (94eabc2) with your PR commit (c4da9ce). Each time is the median of 7 runs, and the "missing pixels" column comes from those same two commits. It's in a code block so the raw markdown survives copying, and it's also saved as pr_perf_table.md in my scratchpad (/tmp/claude-53807/-home-davidh-repos-git-pyresample/2c77a686-8f1c-4596-acbb-2a73545294c7/scratchpad/pr_perf_table.md).

Case	Source outline vertices	main	This PR	Change	Missing pixels (main → PR)
target_inside_full_disk	50 → 50	229 ms	239 ms	1.04×	1 → 1
target_inside_partial_disk	22 → 37	153 ms	178 ms	1.16×	272 → 1
target_crosses_clipped_edge	22 → 37	99 ms ¹	191 ms	1.92×	NotImplementedError → 0
target_crosses_clipped_edge_and_limb	22 → 37	42 ms	75 ms	1.79×	17 → 2
target_crosses_two_clipped_edges	11 → 22	54 ms	101 ms	1.85×	16 → 0
target_crosses_clipped_edge_and_opposite_limb	11 → 22	48 ms	77 ms	1.61×	85 → 0
target_contains_source	11 → 22	43 ms	78 ms	1.82×	0 → 0
target_crosses_limb	50 → 50	196 ms	195 ms	1.00×	2 → 2
target_contains_full_disk	50 → 50	343 ms	343 ms	1.00×	5 → 5
target_is_other_partial_disk ²	22 → 37	114 ms	370 ms	3.26×	319 → 16
RSS → euro4, 1024×1024 (#728)	22 → 37	305 ms	337 ms	1.10×	273 → 1
RSS → eurol, 2560×2048	22 → 37	386 ms	863 ms	2.24×	549 → 0
Total		2012 ms	3048 ms	1.51×

Times are the median of 7 calls to AreaDefinition.get_area_slices with slice caching off, comparing main (94eabc2) and this PR (c4da9ce) on the same machine. The first ten rows are the cases in the new test_get_area_slices_geos_cross_projection_coverage test. "Missing pixels" is the largest number of needed source rows or columns cut off on any side, where "needed" means hit by at least one target pixel center.

¹ Time until main raises NotImplementedError.
² The target is also a partial geostationary disk, so both outlines gain vertices.

Full-disk sources are unchanged because their outline is identical. Partial-disk sources are 1.1–3.3× slower. Nearly all of the time is in the pure-Python spherical polygon intersection, which slows down as vertices are added.

• Largest slowdown, 3.26×: target_is_other_partial_disk, because the target is also a partial disk and both outlines gain vertices.
• Real satpy-size cases: RSS → euro4 (gradient_search leaves part of the destination empty for a partial geostationary disk #728) takes 0.3 s on both main and the PR. RSS → eurol goes from 0.39 s to 0.86 s.

[djhoese]

I've now rebased this branch and moved the work in #729 to use the same functionality that was used in this PR. Pretty much the same logic just a little more flexible. Now the two slicing code branches use the same code for geostationary segmenting.

[djhoese]

Here's an ABI CONUS case that Claude remapped using nearest neighbor to a lon/lat grid.