Mirror scanpy's image-comparison tolerances to fix CPU plotting test failures

[sueoglu]

Mirror scanpy's image-comparison tolerances to fix CPU plotting test failures

Fixes #1082

CPU test suite (hatch-test.py3.12, hatch-test.py3.14) was failing on
image-comparison plotting tests with AssertionError. The comparisons go through check_same_image in tests/conftest.py, which wraps
matplotlib.testing.compare.compare_images(expected, actual, tol=tol) The tol is an RMS pixel difference on a 0–255 scale. ehrapy was setting it far too strictly (1e-1–2e-1 ≈ near pixel-perfect), whereas scanpy uses the same mechanism with tol of 5–40 (most commonly 15). matplotlib is pinned and resolves transitively to 3.10.5; small rendering differences across environments blow past the tiny threshold even when plots are visually equivalent.

What this PR does

Mirrors scanpy's tolerance approach for the tolerance-solvable failures (30
tests):

• temporary mirror of matplotlib in the scanpy dependency list (pyproject.toml): once released scanpy carries the exclusion, it can be removed

• tests/conftest.py: check_same_image now defaults to tol=15 (scanpy's
  most common value); individual tests bump it only where rendering legitimately
  differs more.

• Per-test tolerances assigned in clean tiers (each ≥1.3× the observed RMS):

  tier	used for
  15 (default)	sankey, cohort flowchart, scanpy scatter/violin/clustermap/pca/pca_overview
  25	missingno (all), scanpy heatmap/dotplot/tracks/matrix/stacked_violin/rank_heatmap/rank_dotplot/pca_variance
  35	cohort barplots, scanpy dpt_timeseries
  40	scanpy pca_loadings (highest RMS, 30.8)

• Sensitivity tests keep their tight tol=1e-1 (e.g.
  test_CohortTracker_plot_cohort_barplot_test_sensitivity,
  test_CohortTracker_flowchart_image_sensitivity) so they still detect
  intentional differences — the tiers were chosen to remain meaningful, not so
  loose they defeat the comparison.

Two failures are not tolerance-fixable:

• test_catplot_vanilla, test_stratified_table_one_plot — image size
  mismatch (tolerance-independent; compare_images bails before computing RMS) -> pinned matplotlib

Reference code on scanpy repo for image comparison tolerance: https://github.com/scverse/scanpy/blob/ee7707bc208132cca8387e542d0532f6967f68cc/tests/test_plotting.py#L567

Reference on scanpy pyproject.toml for pinning matplotlib: https://github.com/scverse/scanpy/blob/d5a745e1258f5ca6a23c6a790f27b40d7ad05418/pyproject.toml#L60

[eroell]

• In the PR comment, please use permalinks. This is generally advisable when referring to code. This link currently used will be outdated at one point: https://github.com/scverse/scanpy/blob/main/tests/test_plotting.py#L567. But if you copy a permalink, it remains clear what was referred to, even if scanpy's coded changes: https://github.com/scverse/scanpy/blob/ee7707bc208132cca8387e542d0532f6967f68cc/tests/test_plotting.py#L567
• Linking the scanpy script for plot testing as I suggest in a comment gives indeed quite an immediate glance at how scanpy sets the testing thresholds to 15-40, depending on the plot; much more loose than we currently had it
• Sorry I made accidentally a few comments standalone, should have all been in one review pass to keep the inbox messages at bay