geotiff tests: 8 failures from stale patch target and pre-multiple-of-16 tile_size value

[brendancol]

Describe the bug

Running pytest xrspatial/geotiff/tests produces 8 failures, all in two files:

test_predictor2_big_endian_gpu_1517.py (7 failures):

AttributeError: <module 'xrspatial.geotiff' ...> has no attribute 'read_to_array'

_block_cpu_fallback does monkeypatch.setattr(geotiff_pkg, 'read_to_array', _no_fallback, raising=True). After PR #1885 moved read_geotiff_gpu into xrspatial/geotiff/_backends/gpu.py, the package no longer exposes read_to_array at that name — __init__.py imports it as _read_to_array, and the actual call site is the _read_to_array symbol inside _backends/gpu.py. The monkeypatch with raising=True therefore aborts the test before the GPU path runs.

test_size_param_validation_gpu_vrt_1776.py::test_tile_size_positive_works:

ValueError: tile_size must be a positive multiple of 16 (TIFF 6 spec
requirement for TileWidth/TileLength), got tile_size=4; try tile_size=16.

The test passes tile_size=4. The validator now requires multiples of 16 (a later change), so the original value is no longer valid. The test's intent (small but valid tile size still works) is fine; only the literal needs updating.

Expected behavior

pytest xrspatial/geotiff/tests should run clean.

Fix

• Patch the call site xrspatial.geotiff._backends.gpu._read_to_array instead of the no-longer-existing xrspatial.geotiff.read_to_array.
• Change the tile_size=4 assertion to tile_size=16, the smallest currently valid value, and update the docstring accordingly.

Additional context

CPU and dask tests are unaffected. The failures appear only on GPU-enabled environments (these tests are gated on cupy + CUDA + tifffile).