geotiff: GPU stripped windowed coords honour has_georef=False (#1753)

[brendancol]

Summary

Fix backend inconsistency on open_geotiff(path, gpu=True, window=...) for non-georef stripped TIFFs. PR #1738 (the GPU stripped fallback that forwards max_pixels / window / band to the CPU helper) regressed the #1710 fix for the GPU stripped fallback path: a non-georef file read with gpu=True, window=... returned float64 coords synthesised from the default unit GeoTransform (e.g. y=[-0.5, -1.5, -2.5, -3.5]), while the eager numpy, dask, and tiled GPU paths returned integer pixel coords (np.arange(c0, c1, dtype=int64)).

The fix routes the has_georef=False case through the integer-coord branch on the GPU stripped fallback, mirroring the eager and dask paths.

Root cause

The windowed coord branch added in #1738 only checks if t is None:. That is unreachable because _extract_transform always returns a default GeoTransform() instance with has_georef=False. The default transform has pixel_width=1.0, pixel_height=-1.0, so the PixelIsArea sub-branch produced (c0 + 0.5) * 1.0 + 0 and (r0 + 0.5) * -1.0 + 0 -> [-0.5, -1.5, ...].

dask+cupy shares the same code path, so the bug fired on every gpu=True windowed read of a non-georef stripped TIFF regardless of chunking.

Fix

• Route not getattr(geo_info, 'has_georef', True) through the integer-coord branch alongside the existing t is None check at __init__.py:2724-2747.
• Mirror the eager numpy path's windowed coord shortcut at open_geotiff lines 834-836.
• Also fix a latent off-by-one in the integer-coord branch: was arange(r1-r0) instead of arange(r0, r1), so an offset window like (2, 3, 6, 7) would have returned [0, 1, 2, 3] instead of [2, 3, 4, 5]. The branch was unreachable so the off-by-one was latent.

Tests

14 new regression tests in test_gpu_stripped_no_georef_window_1753.py:

• TestStrippedGpuWindowedNoGeoref: origin window, offset window, dask+cupy variants, uint16 dtype, no transform attr
• TestStrippedGpuWindowedBackendParity: dtype-parity (3 windows) and value-parity (3 windows) across all four backends
• TestStrippedGpuWindowedGeorefStillWorks: PixelIsArea origin and offset windows still get float64 coords with the correct half-pixel shift

Test plan

• All 14 new regression tests pass.
• All 15 existing tests in test_no_georef_windowed_coords_1710.py pass (3 previously failed in TestGpuWindowedCoords and TestBackendParity).
• All 2005 non-stale geotiff tests still pass.
• 10 pre-existing failures in test_predictor2_big_endian_gpu_1517.py are unrelated stale test-infrastructure from PR geotiff: read_to_array leaks into public namespace but is not in __all__ or docs #1708 (renamed read_to_array to _read_to_array).

Closes #1753