EPIC: GPU-decoding from TIFF -> CUDA mem -> dlpack -> CuPy

[weiji14]

Working towards fully GPU-native decompression of (Geo)TIFFs into device memory (CUDA GPU) on Rust/FFI side, and zero-copy to CuPy arrays on Python side.

Steps:

• TIFF to CUDA device memory via nvtiff - ✨ Decode GeoTIFF images to CUDA stream #27
• CUDA memory to DLPack in-memory format via dlpark - 💥 Read into CUDA memory DLPack tensor #57
• DLPack to CuPy via raw pointers - ✨ Python bindings for CudaCogReader #58

Notes:

• Expose this through a read_geotiff_dlpack API function?
• Bump min numpy version to 2.2 to get MAINT: from_dlpack thread safety fixes (#28883) numpy/numpy#28889
• Consider pathway through cuTENSOR -> DLPack? CuPy uses cuTENSOR apparently, but unsure if ABI is stable.
• Potentially bring this to cupy-xarray in the future

References:

• https://crates.io/crates/nvtiff-sys
• https://dmlc.github.io/dlpack/latest/
• Array Ecosystem Interoperability zarrs/zarrs#113
• Creating a cupy device array from GPU Pointer cupy/cupy#4644