resample: fix dask nearest/bilinear chunk-seam values on downsampling (#2610)

[brendancol]

Closes #2610

What

• The interp dask path used a fixed overlap depth of 1 input pixel for nearest and bilinear. Block-centered mapping sends an output pixel to input coordinate (o + 0.5) * (in/out) - 0.5, which on a downsample can land up to ~(in/out)/2 pixels past the chunk _output_chunks assigned it to. With depth 1 the overlapped block did not contain that source row/column, so map_coordinates(mode='nearest') clamped to the block edge and returned wrong values for whole chunk-seam rows.
• Adds a per-axis downsample radius (_downsample_radius) to the interp overlap depth in _run_dask_numpy and _run_dask_cupy, clamped per axis exactly as before. The aggregate path already accounted for this; the interp path now matches.

Backends

• dask+numpy and dask+cupy: fixed.
• numpy and cupy (eager): unaffected, no chunking involved.

Test plan

• New TestInterpDaskDownsampleSeam covers nearest/bilinear/cubic at ratios 0.33/0.2/0.1 with chunk layouts that do not divide the input evenly, plus an asymmetric-per-axis case. Fails on main, passes here.
• Full test_resample.py suite: 217 passed (includes cupy and dask+cupy parity on a CUDA host).
• Off-band sweep of 660 ratio/size/chunk/method combinations: dask matches eager numpy within float32 round-off.

[brendancol]

PR Review: resample dask nearest/bilinear chunk-seam fix (#2610)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• _downsample_radius re-imports math at function scope. _ensure_min_chunksize and the aggregate branches already do the same local import, so this matches existing style. A module-level import math would be cleaner, but that pattern predates this PR, so leave it for a bug fix.

What looks good

• Root cause is correctly identified, and the fix mirrors the depth logic the aggregate path already uses. _downsample_radius returns 0 for upsampling, so the upsample paths keep depth 1 and pay nothing.
• The augmented depth feeds the min_size chunk-sizing, the _add_overlap call, and the block function's origin offset from the same value, so they cannot drift apart.
• The per-axis clamp against global_in - 1 is preserved, so tiny inputs and skinny axes still work. Checked on 2x2, single-chunk, and skinny-column cases.
• Both _run_dask_numpy and _run_dask_cupy are patched, so dask+numpy and dask+cupy stay in parity. Eager numpy/cupy are untouched.
• Test coverage holds up: the new class fails on main and passes here, uses random data (a gradient would hide the bug), and uses chunk sizes that do not divide the input evenly. An off-band sweep of 660 ratio/size/chunk/method combinations all matched eager numpy within float32 round-off.

Checklist

• Algorithm matches the block-centered mapping; depth now covers the downsample displacement
• All four backends consistent (cupy and dask+cupy verified on a CUDA host)
• NaN handling unchanged (fix only widens overlap)
• Edge cases covered (small inputs, skinny axes, asymmetric per-axis ratios)
• Dask chunk boundaries handled correctly (the point of the fix)
• No premature materialization or extra copies
• Benchmark not needed (bug fix, no new API)
• README feature matrix already lists resample
• Docstrings present; new helper documents the radius rationale