Security: OOB GPU read when TIFF tile grid exceeds TileOffsets length

[brendancol]

Describe the bug

_assemble_tiles_kernel in xrspatial/geotiff/_gpu_decode.py reads tile_out_offsets[tile_idx] where tile_idx is computed from the output pixel's position (tile_row * tiles_across + tile_col). If a TIFF declares image dimensions implying more tiles than its TileOffsets tag supplies, tile_idx can exceed len(tile_out_offsets). The kernel then reads past the end of device memory.

The CPU path (_read_tiles) silently skips tiles whose index exceeds the offsets array, so the caller gets a zero-padded output instead of an error.

Reproduction

Craft a TIFF where:

• ImageWidth / ImageLength and TileWidth / TileLength imply tiles_across * tiles_down = N tiles.
• TileOffsets tag has count M with M < N.

On the GPU path (open_geotiff_gpu), the kernel launches one thread per output pixel. Threads whose pixel maps to a tile index >= M read past the end of d_decomp_offsets.

Expected behavior

The reader should reject the file with a ValueError during header parsing, naming the mismatch between the declared tile grid and the TileOffsets length.

Impact

• GPU path: out-of-bounds device read. Produces garbage pixels, or trips a CUDA illegal memory access that kills the process.
• CPU path: silent wrong output (zero-filled tiles where offsets were missing).
• Adversarial input only. Files written by xrspatial.geotiff._writer or GDAL always have matching counts.

Context

Found during /sweep-security audit of the geotiff subpackage. Related to #1215 (tile dimension guard) and a separate predictor-kernel finding filed as a sibling issue.

Desktop:

• OS: Linux
• Python: 3.11
• xarray-spatial: current master