set_overlay_mask was unusable in tile mode — no mask shape could be drawn

[CSSFrancis]

In tile mode there was no mask shape that worked. The renderer sizes the
overlay against base_width || image_width — the OVERVIEW grid — but
enable_tile sets image_width/image_height to the FULL native frame, and
the shape check compared only against those:

• an overview-sized mask (what the renderer wants) raised ValueError here;
• a full-resolution mask was accepted, encoded, and then silently discarded
  by the renderer — mBytes.length===iw*ih else maskCache=null. No error, no
  overlay, nothing in any log.

A caller could not get this right, because there was no right answer.

Measured, on a 4096×4096 tiled plot

tile_enabled=True  image=(4096x4096)  base=(1024x1024)

BEFORE
  overview-sized (what the RENDERER wants) → REJECTED by python:
      ValueError: mask shape (1024, 1024) does not match image (4096 x 4096)
  full-sized     (what PYTHON validates)   → ACCEPTED, 22.4 MB of base64,
      renderer expects 1048576 bytes, got 16777216 → DROPPED SILENTLY

AFTER
  overview-sized → ACCEPTED, ships 1048576 bytes → MATCH
  full-sized     → ACCEPTED, ships 1048576 bytes → MATCH

Both shapes are accepted now, and a full-resolution mask is reduced to the
overview grid before encoding — so the same call works tiled or not, and ships
1.4 MB instead of 22.4 MB.

Why block ANY and not a subsample

A mask marks where objects are. Objects a few pixels across must not vanish
because the sampled pixel missed them, and striding 4096 → 1024 keeps one pixel
in sixteen. _reduce_mask_any folds the uneven tail rows/columns into the last
block rather than cropping them, so an object at the right or bottom edge
survives too.

What it looks like

Found via SpyDE, whose segmentation overlay is a mask on a tiled 4096² frame.
Before, the overlay simply never appeared — the preview reports thousands of
instances and the figure shows none of them:

before — mask silently dropped	after — mask draws

Tests

Two, both asserting the bytes that actually ship rather than the array
handed in — which is the shape of assertion that would have caught this:

• either input shape ships exactly base_width*base_height bytes and the
  reduced mask is non-empty;
• a single lit pixel deliberately off a block boundary survives the reduction
  (a subsample would drop it).

2002 passed, 58 skipped on the full suite.

Note for downstream

SpyDE's _set_raster_overlay did its own overview reduction to work around
this and got the ValueError for its trouble. It now hands over the native
mask and lets this own the arithmetic, so there is one owner and it cannot
disagree with the shape check beside it. That companion change is in
CSSFrancis/spyde feat/drift-particles.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 85.71429% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 90.74%. Comparing base (9784741) to head (a61d865).

Files with missing lines	Patch %	Lines
anyplotlib/plot2d/_plot2d.py	85.71%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #56      +/-   ##
==========================================
- Coverage   90.77%   90.74%   -0.04%     
==========================================
  Files          40       40              
  Lines        4500     4526      +26     
==========================================
+ Hits         4085     4107      +22     
- Misses        415      419       +4     

☔ 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.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.