fix(downsample): shift output origin to coarse-grid pixel centers

[thewtex]

Summary

Fixes the output-origin defect in the downsample and downsample-label-image pipelines reported in #1409. Both pipelines resampled onto a coarser grid but kept the input origin unchanged, so the downsampled image was misaligned by half a coarse pixel relative to the physically correct grid — and relative to downsample-bin-shrink, which was already correct.

The problem

Under ITK's pixel-center convention, a pixel's origin is the physical location of its center. When you downsample by a factor k, each output pixel covers k input pixels, so its center sits at the physical location of continuous input index (k − 1) / 2. The old code ignored this:

outputOrigin[i] = inputOrigin[i] + cropRadiusValue * inputSpacing[i];

This kept the origin at the first input pixel's center (plus any crop offset) with no half-pixel shift and no awareness of the image direction matrix. itk::BinShrinkImageFilter (used by downsample-bin-shrink) already does the right thing, which is why the two pipelines disagreed.

The fix

At all three affected sites — downsample.cxx scalar path, downsample.cxx vector-image path, and downsample-label-image.cxx — the origin is now computed through a continuous index so ITK applies the direction matrix for us:

// Origin lands at the center of the first bin, matching itk::BinShrinkImageFilter
outputOriginIndex[i] = inputStartIndex[i] + cropRadiusValue + (shrinkFactors[i] - 1.0) / 2.0;
...
inputImage->TransformContinuousIndexToPhysicalPoint(outputOriginIndex, outputOrigin);

Key points:

• Half-pixel shift. The (shrinkFactors[i] - 1.0) / 2.0 term moves the origin to the center of the first coarse bin — the (k − 1) / 2 convention BinShrinkImageFilter uses.
• Direction-aware. Building an itk::ContinuousIndex and calling TransformContinuousIndexToPhysicalPoint routes the shift through the image direction matrix, instead of the old per-axis arithmetic that silently assumed an identity direction.
• Crop radius composes correctly. The existing crop offset is folded into the same continuous index, so cropRadius and the origin shift now stack (e.g. cropRadius = [8, 8] with shrinkFactors = [2, 2] → origin [8.5, 8.5]).
• downsample-bin-shrink.cxx was already correct and is unchanged; the fix makes downsample match it exactly for the same inputs.

Verification & tests

Added regression tests (Node + Python, WASI and emscripten) that pin the corrected grid convention so it cannot silently regress:

• downsample(cthead1, shrinkFactors=[2, 2]) → origin [0.5, 0.5], spacing [2, 2], size [128, 128]
• downsample origin/spacing/size now exactly equal downsample-bin-shrink for the same input and factors (directly pins the downsample pipeline should result in a shift in origin #1409 convention)
• cropRadius = [8, 8] composes with the shift → origin [8.5, 8.5], size [120, 120]
• vector-image path asserts the half-spacing shift per axis

Baselines were regenerated for the corrected origins and the test-data hash bumped accordingly. The package version is bumped 1.8.1 → 1.9.0 (behavior change to the output geometry).

Fixes #1409