Changes in version 0.1.0                        

  - Add Windows CUDA support via pre-compiled companion DLL loaded
    dynamically through sfgpu::sfgpu_install_cuda().
  - Add sfgpu_within_distance() for bounded-memory, sparse point matches
    within an inclusive Euclidean radius on CPU and available CUDA or
    Metal backends.
  - Accept coordinate-unit numeric radii for matrices and spatial
    inputs, plus convertible units radii for projected spatial inputs.
  - Use sorted-y indexes, device binary search, and candidate compaction
    for CUDA and Metal radius calls when the index workspace fits
    tile_bytes; either GPU otherwise uses its bounded tiled scan. Metal
    compares ordered binary64 keys and caps each compaction command
    at 65,536 candidates. This adds no performance guarantee.