feat: GPU acceleration module and scripts

[FIrgolitsch]

Stacked PR 2/22 — review order: #115 → #97 → #98 → #99 → #100 → #101 → #108 → #106 → #107 → #87 → #116 → #110 → #111 → #40 → #112 → #113 → #117 → #118 → #120 → #121 → #122 → #123 → #124 → #125

Base: pr-e-motor-stacking. Retargets to main as upstream PRs merge.

---

PR #99 — GPU Acceleration (unified CPU/GPU scripts)

Adds the linumpy/gpu/ module (CuPy + JAX primitives) and wires the existing CPU scripts in the repo to pick up an optional GPU path via --use_gpu. Stand-alone *_gpu.py script variants are intentionally removed in this PR (replaced by the unified CPU scripts in #97).

New library module linumpy/gpu/

• array_ops.py — cupy/jax array primitives and CPU fallbacks
• corrections.py — GPU illumination / flat-field corrections
• cuda_env.py — CUDA environment detection (CUDA version, CuPy availability, JAX plugin status) + device selection; used by every GPU-capable script to negotiate the backend
• fft_ops.py — GPU FFT (phase-correlation building blocks)
• image_quality.py — GPU variance / sharpness metrics
• interpolation.py — GPU affine/linear resampling
• morphology.py — GPU morphological ops (erode/dilate/tophat)
• normalization.py — GPU intensity normalization (percentile / rolling-ball) paired with linumpy/preproc/normalization.py from feat: utility modules, preprocessing improvements, galvo correction #97
• registration.py — GPU phase-correlation + refinement; callable from linumpy/stitching/registration.py

Scripts

• scripts/linum_benchmark_gpu.py — benchmark harness comparing CPU vs GPU on the key hot paths (normalization, registration, resampling)
• scripts/linum_gpu_info.py — report available CUDA / CuPy / JAX devices and plugin status

Removed (unified into CPU scripts via --use_gpu)

• linum_aip_gpu.py, linum_assess_slice_quality_gpu.py, linum_create_mosaic_grid_3d_gpu.py, linum_estimate_transform_gpu.py, linum_fix_illumination_3d_gpu.py, linum_generate_mosaic_aips_gpu.py, linum_normalize_intensities_per_slice_gpu.py, linum_resample_mosaic_grid_gpu.py

Shell helpers

• shell_scripts/fix_jax_cuda_plugin.sh — resets a broken JAX CUDA plugin install; autodetects CUDA 12 vs 13 and picks the matching jax-cuda*-plugin wheel

Notes

• build: migrate to uv, add ruff/ty, format all files #95 declares the gpu / gpu-cuda12 / gpu-cuda13 optional extras used here
• Unified CPU scripts live in feat: utility modules, preprocessing improvements, galvo correction #97 (and feat: diagnostic and analysis scripts #100 for diagnostics); those accept --use_gpu and dispatch into linumpy/gpu/