Fix dask aggregate boundary contamination and clean up cumulative bookkeeping

[brendancol]

Summary

Fixes #1469.

• Aggregate dask overlap now uses boundary=np.nan (cupy.nan in the cupy mirror). The aggregate kernels already skip NaN inputs and return NaN for empty windows, so the pad is ignored. The existing _agg_block_np indexing happens to keep output windows out of the global-edge pad, but the NaN boundary makes that contract explicit instead of relying on it.
• Compute the scale-driven and depth-driven min_size together, call _ensure_min_chunksize once, then build the cumulative arrays once. Removes the wasted first compute and the recompute branch that used data.chunks[0] != tuple(cum_in_y[1:] - cum_in_y[:-1]) as a roundabout chunk-equality check.
• Interp dask paths stay on boundary='nearest' so they keep matching the scipy mode='nearest' semantics that the eager numpy interp path uses.

Test plan

• pytest xrspatial/tests/test_resample.py passes (68 tests: 62 existing + 6 new).
• New TestAggregateDaskBoundary pins dask aggregate min/max/median to bit-equal eager numpy:
  • test_chunk_spanning_window_bit_identical: chunks (7, 7) on a 24x24 input force output windows to span chunk boundaries.
  • test_global_edge_extremes_match_eager: 999 and -999 on every outer row/column.
• Existing aggregate-parity tests still pass for average, min, max, median, mode.
• Interp dask parity (nearest, bilinear, cubic) unchanged.

Note on the boundary change

The _agg_block_np indexing reads from block[depth_y : depth_y + chunk_size, ...], which never touches the global-edge pad. So boundary='nearest' was not actually corrupting results in the current code, and the new tests would not have failed pre-fix. The change is still worth making:

1. It removes a fragile coupling between kernel indexing and the overlap padding choice. If someone later extends _agg_block_np to walk further into the block, NaN padding fails safely while 'nearest' silently biases min/max/median.
2. It matches how every other map_overlap call site in xrspatial sets its boundary.

The bookkeeping cleanup is the substantive fix; the boundary change is defence-in-depth.