💥 Support reading into CuPy arrays via cog3pio xarray backend

[weiji14]

Proof of concept interface to read TIFF data into an xarray.DataArray via Cog3pioBackendEntrypoint. Handles reshaping of the 1-D array returned from CudaCogReader into a proper 3-D array of shape (Channels, Height, Width).

Preview at https://cog3pio--71.org.readthedocs.build/en/71/api/#cog3pio.xarray_backend.Cog3pioBackendEntrypoint

Example usage:

import cog3pio
import cupy as cp

da: xr.DataArray = xr.open_dataarray(filename_or_obj="TCI.tif", engine="cog3pio", device=(2, 0))
assert isinstance(da, cp.ndarray)

TODO:

• Implement xy_coords method on PyCudaCogReader struct (cherry-picked to ✨ Implement xy_coords method on PyCudaCogReader struct #82)
• Initial attempt on getting xarray.open_dataarray(..., engine="cog3pio") to read TIFF data into CuPy-backed arrays in xarray.
• Refactor into a more intuitive API, consider putting into cupy-xarray? Yep - see TIFF to GPU memory via cog3pio backend entrypoint xarray-contrib/cupy-xarray#81
• Document example usage
• Update main README.md roadmap

Notes:

• The use_cuda flag doesn't feel very ergonomic or future proof. Am thinking that there could be some way to conform better to the array API standard, by letting users to set a device which is passed to the from_dlpack() method somehow edit: no, device should be set when DLPack object is instantiated. This means users can control where the TIFF data is decoded, be it CPU or a CUDA GPU (with a specific GPU id?)
• Want to have a way to default to GPU-based decoding somehow, with a fallback to CPU. Also should default to per-thread default stream instead of legacy default stream.

References:

• https://data-apis.org/array-api/2024.12/design_topics/device_support.html

Closes #26

[codspeed-hq]

Merging this PR will not alter performance

✅ 7 untouched benchmarks

---

_{Comparing xarray_cudacog (60f5605) with main (bd9b65a)}

[weiji14]

Thinking that the Cog3pioBackendEntrypoint code should just live in cupy-xarray, so opened xarray-contrib/cupy-xarray#70. Considering to eventually deprecate xarray as a dependency in this repo.