Decode TIFF predictor=2 sample-wise for multi-byte integer round-trip

[brendancol]

Summary

• TIFF predictor=2 decode and encode were byte-wise at a sample stride, which loses carries inside multi-byte samples. Files xrspatial wrote round-tripped because both halves were wrong in matching ways, but uint16, int16, uint32 and int32 files written by libtiff/GDAL with predictor=2+deflate/lzw silently decoded to wrong pixel values whenever an adjacent-sample difference crossed a byte boundary.
• Decode now operates on whole sample values in the file's byte order, per-channel for chunky multi-band data; the 8-bit fast path stays. Predictor=2 on floats stays byte-wise (the spec restricts predictor=2 to integers; floats are out of spec for this predictor) so old xrspatial-written float files still read back unchanged.
• The GPU encode dispatch was also fixed: it had been called with row_bytes multiplied by samples-per-pixel twice, which was inert only because the byte-wise math was wrong on both sides.

Repro before the fix

import rasterio, numpy as np
arr = np.array([[1000, 2000, 3000, 4000]], dtype='uint16')
with rasterio.open('/tmp/r.tif', 'w', driver='GTiff',
                   height=1, width=4, count=1, dtype='uint16',
                   compress='deflate', predictor=2) as dst:
    dst.write(arr, 1)

from xrspatial.geotiff._reader import read_to_array
print(read_to_array('/tmp/r.tif')[0])  # before: [1000 1744 2488 3232]

Test plan

• New test_predictor2_samplewise.py: rasterio (libtiff) round-trips for LE uint16/int16/uint32/int32 with carries
• tifffile round-trips for big-endian uint16/int16/uint32/int32
• Multi-band chunky uint16 round-trip against libtiff
• xrspatial-written uint16 readable by rasterio (interop the other direction)
• GPU multi-byte decode parity for libtiff predictor=2 files
• Existing predictor and writer test suites still pass (CPU and GPU paths)

[brendancol]

Closing as duplicate of #1498 (already merged). The audit agent that opened this PR was working from a stale local main that didn't have #1498 yet, so it re-derived the same fix. Verified: with #1498 applied, the repro [1000, 2000, 3000, 4000] uint16 predictor=2 round-trips correctly through rasterio.