geotiff: cover degenerate-shape reads on dask and GPU backends

[brendancol]

Summary

The eager numpy reader already covers degenerate-shape sources (1x1, 1xN, Nx1), and the dask streaming write path covers writing degenerate dask rasters. What was missing is the read side on the non-eager backends:

• the windowed dask reader (chunks=) splitting a single-pixel dimension into chunks,
• the GPU reader (gpu=True) launching its decode kernels on a degenerate grid,
• and the dask+gpu combination.

The smallest dask-read source in the existing suite is 8x8 / 2x32, and the parity-matrix fixtures are 32x32 / 64x64, so these shapes never reached the windowed read or the GPU launch.

I probed all three backends on these shapes first. They read correctly today, so this is a coverage gap rather than a bug, and no source code changes.

What was added

xrspatial/geotiff/tests/read/test_degenerate_shapes.py, 18 tests:

• dask read across chunk sizes that meet, split, and exceed the degenerate dimension, each asserting pixel parity against the eager read
• dask read coord / transform / CRS preservation
• GPU read and dask+gpu read, both checked against the eager read

The fixture passes an explicit attrs['transform'] because the writer cannot infer a pixel size from a single-element coord axis.

Validation

All 18 tests pass on a CUDA host. The 6 GPU and dask+gpu cells ran rather than skipped, and the 1x1 GPU read goes through a grid-size-1 kernel launch. flake8 clean.

Found by the test-coverage deep-sweep against the geotiff module.