Fill sparse COG tiles with nodata instead of crashing

[brendancol]

Summary

GDAL's SPARSE_OK=TRUE writes blocks consisting entirely of nodata with TileByteCounts (or StripByteCounts) == 0 and a matching offset of 0. The reader handled this badly:

• Local mmap path tried to decompress the empty byte range and raised error -5 while decompressing data: incomplete or truncated stream.
• COG HTTP path skipped the empty fetch but left the result as np.empty, so sparse regions returned uninitialised memory.
• GPU path crashed inside the decode pipeline.

This branch detects sparse blocks before decode, pre-fills the result with the file's GDAL_NODATA value (zero when no nodata tag is present, matching GDAL's convention), and skips decode for those entries. The GPU dispatcher routes sparse files through the CPU reader and copies the result onto the device, since the GPU pipeline does not handle empty ranges.

Repro on main

import rasterio, numpy as np
from xrspatial.geotiff import open_geotiff

data = np.zeros((128, 128), dtype=np.uint16)
data[:64, :64] = 100
profile = {
    'driver': 'GTiff', 'dtype': 'uint16',
    'height': 128, 'width': 128, 'count': 1,
    'tiled': True, 'blockxsize': 64, 'blockysize': 64,
    'compress': 'DEFLATE', 'SPARSE_OK': 'TRUE', 'nodata': 0,
}
with rasterio.open('/tmp/sparse.tif', 'w', **profile) as dst:
    dst.write(data[:64, :64], 1, window=rasterio.windows.Window(0, 0, 64, 64))

open_geotiff('/tmp/sparse.tif')
# OSError: error -5 while decompressing data: incomplete or truncated stream

After the fix the call returns a (128, 128) float64 DataArray with the filled tile and NaN over the sparse region.

Test plan

• pytest xrspatial/geotiff/tests/test_sparse_cog.py (5 new tests covering tiled/stripped, with/without nodata, raw + accessor reads, and GPU)
• pytest xrspatial/geotiff/tests/ full suite (651 passed, 7 deselected — 3 pre-existing matplotlib palette RecursionError failures unrelated to this PR, also fail on origin/main)
• Manual repro against GDAL SPARSE_OK=TRUE files for tiled and stripped layouts and CuPy GPU read