Add GPU acceleration subpackage linumpy.gpu (CuPy + CPU fallback)

[FIrgolitsch]

Summary

Adds a new linumpy.gpu subpackage: GPU-accelerated implementations of the compute-heavy operations used elsewhere in the library (FFT phase correlation, affine interpolation, morphology, B-spline fitting, bias-field correction, image-quality metrics, tissue-interface detection, and Zarr I/O), all with automatic fallback to NumPy when CuPy/CUDA is unavailable.

Land after #139.

What changed

New package linumpy/gpu/ (13 modules, all new):

Module	Provides
__init__.py	Device detection, multi-GPU selection (file-locked round-robin across physical devices, honouring CUDA_VISIBLE_DEVICES), to_gpu/to_cpu/get_array_module helpers, list_gpus/select_gpu/select_best_gpu, env-gated via LINUMPY_USE_GPU
interface.py	GPU tissue-interface detection (mirrors linumpy.geometry.interface)
array_ops.py	GPU array utilities
fft_ops.py	GPU phase correlation / FFT ops
interpolation.py	GPU affine transform + resampling
morphology.py	GPU morphological filters
corrections.py	GPU correction kernels
bspline.py	GPU B-spline fit/evaluate
bias_field.py	GPU bias-field correction
image_quality.py	GPU image-quality metrics
zarr_io.py	GPU-aware Zarr I/O
kvikio_zarr.py	GPUDirect-Storage (kvikio) Zarr reader
nvcomp_zstd.py	GPU ZSTD compression via nvCOMP

Plus linumpy/tests/test_gpu_bspline.py — B-spline correctness (partition-of-unity, constant/linear recovery, mask handling) and CPU/GPU numerical agreement on the fit + evaluate paths.

Design notes

• Transparent CPU fallback. Every public entry point returns plain np.ndarray / ArrayLike; callers stay device-agnostic. CuPy is imported lazily and gated on availability, so importing the package never requires CUDA.
• Device selection. On multi-GPU hosts a file-locked counter round-robins across devices to avoid the race where concurrent forks all grab the same "most-free-memory" snapshot. Single-GPU or externally-pinned (CUDA_VISIBLE_DEVICES) hosts short-circuit to device 0.
• Typing. Optional GPU deps are typed as Any via the [tool.ty] replace-imports-with-any list (cupy, cupyx, numba, kvikio, nvidia), so the type checker stays green without CUDA installed.

Reviewer focus

• The CPU/GPU agreement in test_gpu_bspline.py is the main automated guard for numerical equivalence — worth a close read since real CUDA isn't exercised in CI.
• Device selection (__init__.py) uses a fcntl-locked counter file in $TMPDIR; check that the locking/unlocking and the single-GPU fast path match your deployment (e.g. Nextflow setting CUDA_VISIBLE_DEVICES per task).
• The kvikio/nvcomp paths are CUDA-only and not covered by CI; they're verified manually on the GPU server.

How to verify

uv run ruff check linumpy/gpu/
uv run ty check linumpy/gpu/
uv run pytest linumpy/tests/test_gpu_bspline.py -x -v   # runs the CPU path without CUDA