Summary

• Adds resample() for changing raster cell size without reprojection. Previously the only option was calling reproject() with identical source/target CRS, which runs the full coordinate transform pipeline for no reason.
• 8 methods: nearest, bilinear, cubic (interpolation, up or down) and average, min, max, median, mode (block aggregation, downsampling only).
• Nodata pixels don't contaminate neighbors during interpolation (zero-fill/weight-mask trick). Aggregation methods skip NaN.
• Works on all four backends: numpy, cupy, dask+numpy, dask+cupy. The dask path maps global coordinates per-block so results don't depend on chunk layout.
• resample(da, scale_factor=0.5) or resample(da, target_resolution=30.0). Tuples work for asymmetric scaling.

Test plan

• 51 tests: API validation, output geometry, known-value correctness, NaN handling, edge cases (single pixel, single row, non-square, very small scale), Dataset input
• Dask parity: all 8 methods at multiple scale factors vs numpy
• CuPy parity: interpolation and aggregation vs numpy
• Manual check that user guide notebook renders correctly

Closes #1152