Resample interpolation uses edge-aligned coordinates instead of block-centered

[brendancol]

Describe the bug

resample() with interpolation methods (nearest, bilinear, cubic) uses scipy.ndimage.zoom's coordinate formula, which maps output pixel o to input pixel o * (N_in - 1) / (N_out - 1). This pins the first output pixel to the first input pixel and the last to the last.

The output coordinate metadata (_new_coords) places pixels at block centers, which is the normal convention for raster data. So there's a spatial mismatch between where the interpolation actually samples and where the coordinates say the data is.

The aggregation path (average, min, max, median, mode) uses block-centered sampling and doesn't have this problem.

Reproducer

With a linear gradient where value equals x-coordinate, a correct resample should produce output values that match their coordinate labels:

import numpy as np
from xrspatial.resample import resample
from xrspatial.tests.general_checks import create_test_raster

data = np.tile(np.arange(4, dtype=np.float32), (4, 1))
agg = create_test_raster(data, attrs={'res': (1.0, 1.0)}, dims=['y', 'x'])

out = resample(agg, scale_factor=0.5, method='bilinear')
print(out.x.values)    # [0.5, 2.5]  -- correct block centers
print(out.values[0])   # [0.0, 3.0]  -- wrong, should be [0.5, 2.5]

The interpolation samples at input positions 0.0 and 3.0 (edge-aligned), but the coordinates say the pixels are at 0.5 and 2.5 (block-centered). Half-pixel shift at the edges.

Expected behavior

Interpolation should sample at positions matching the output coordinate metadata. The correct formula is input_pixel = (out_pixel + 0.5) * (N_in / N_out) - 0.5.

Affected code paths

• _run_numpy -- uses scipy.ndimage.zoom directly
• _run_cupy -- uses cupyx.scipy.ndimage.zoom directly
• _interp_block_np -- dask+numpy path, uses map_coordinates with the zoom formula
• _interp_block_cupy -- dask+cupy path, same

Fix

Replace scipy.ndimage.zoom in the non-dask paths with map_coordinates using the block-centered formula. Update the dask block functions to use the same formula.