Add nvJPEG GPU acceleration for JPEG-compressed GeoTIFFs (#1050) (#1055)

Wire JPEG (TIFF tag 7) into the GPU decode and encode pipelines.
When libnvjpeg.so is available, read_geotiff(gpu=True) and
write_geotiff(gpu=True, compression='jpeg') use nvJPEG for
hardware-accelerated JPEG on GPU. Falls back to Pillow on CPU
when nvJPEG is not installed.

Changes:
- _gpu_decode.py: Add _find_nvjpeg_lib/_get_nvjpeg lazy discovery,
  _try_nvjpeg_batch_decode for GPU reads, _nvjpeg_batch_encode for
  GPU writes. Hook tag 7 into gpu_decode_tiles and gpu_compress_tiles.
- _writer.py: Add 'jpeg' to _compression_tag map. Handle JPEG in
  _write_tiled and _write_stripped (call jpeg_compress directly,
  skip predictor). Validate uint8 dtype and 1/3 band constraint.
- __init__.py: Update docstrings with JPEG option.
- README.md: Add nvJPEG to GPU codecs list and JPEG write example.
- tests/test_jpeg.py: 13 tests covering codec round trips, tag
  wiring, tiled/stripped write-read, public API, and dtype validation.
- tests/test_writer.py: Fix test_unsupported_compression (was using
  'jpeg' as the unsupported example).

Author: brendancol

README.md

@@ -158,12 +158,13 @@ read_geotiff('mosaic.vrt')                           # VRT mosaic (auto-detected
 
 write_geotiff(cupy_array, 'out.tif')                 # auto-detects GPU
 write_geotiff(data, 'out.tif', gpu=True)             # force GPU compress
+write_geotiff(data, 'ortho.tif', compression='jpeg') # JPEG for orthophotos
 write_vrt('mosaic.vrt', ['tile1.tif', 'tile2.tif'])  # generate VRT
 ```
 
 **Compression codecs:** Deflate, LZW (Numba JIT), ZSTD, PackBits, JPEG (Pillow), uncompressed
 
-**GPU codecs:** Deflate and ZSTD via nvCOMP batch API; LZW via Numba CUDA kernels
+**GPU codecs:** Deflate and ZSTD via nvCOMP; LZW via Numba CUDA; JPEG via nvJPEG
 
 **Features:**
 - Tiled, stripped, BigTIFF, multi-band (RGB/RGBA), sub-byte (1/2/4/12-bit)

xrspatial/geotiff/__init__.py

@@ -319,7 +319,10 @@ def write_geotiff(data: xr.DataArray | np.ndarray, path: str, *,
     nodata : float, int, or None
         NoData value.
     compression : str
-        'none', 'deflate', or 'lzw'.
+        'none', 'deflate', 'lzw', 'jpeg', 'packbits', or 'zstd'.
+        JPEG is lossy and only supports uint8 data (1 or 3 bands).
+        With ``gpu=True``, JPEG uses nvJPEG for GPU-accelerated
+        encode/decode when available, falling back to Pillow on CPU.
     tiled : bool
         Use tiled layout (default True).
     tile_size : int
@@ -756,7 +759,8 @@ def write_geotiff_gpu(data, path: str, *,
     nodata : float, int, or None
         NoData value.
     compression : str
-        'zstd' (default, fastest on GPU), 'deflate', or 'none'.
+        'zstd' (default, fastest on GPU), 'deflate', 'jpeg', or 'none'.
+        JPEG uses nvJPEG when available, falling back to Pillow.
     tile_size : int
         Tile size in pixels (default 256).
     predictor : bool