feat(n4): GPU-accelerated N4 bias field correction

[FIrgolitsch]

Stacked PR 11/16 — review order: #115 → #97 → #98 → #99 → #100 → #101 → #108 → #106 → #107 → #87 → #116 → #110 → #111 → #40 → #112 → #113

Replaces #109 — that PR was permanently locked closed by GitHub (its base chain-merged had no common history). Recreated here on the proper base pr-3-slice-interpolation. Branch pr-n4-bias-field is unchanged.

---

PR #109 — N4 Bias Field Correction (CPU + GPU)

Adds N4 bias-field correction with CPU/GPU backends, replacing the previous z-intensity normalization.

Library

• linumpy/intensity/bias_field.py — high-level API
• linumpy/intensity/normalization.py — extended for N4 backend, plus fix(normalize): zero agarose per slice while preserving inter-slice brightness (per-slice subtraction of the agarose-median floor pins background voxels to exactly 0 in the normalized output, restoring pre-n4 visualization behavior; global divisor preserves the 2:1 inter-section brightness invariant)
• linumpy/gpu/{bias_field,bspline,n4}.py — GPU kernels (CuPy)
• Removes linumpy/gpu/normalization.py (superseded)

Scripts

• scripts/linum_correct_bias_field.py — replaces linum_normalize_z_intensity
• scripts/diagnostics/linum_benchmark_n4_gpu.py
• scripts/diagnostics/linum_n4_gpu_visual_compare.py

Tests

• linumpy/tests/test_bias_field*.py, test_gpu_{bspline,n4}.py
• linumpy/tests/test_n4_gpu_{equivalency,perf}.py
• Extended test_intensity_normalization.py with test_normalize_volume_agarose_floor_at_zero

Workflow integration

• workflows/reconst_3d/soct_3d_reconst.nf: adds correct_bias_field process
• workflows/reconst_3d/nextflow.config: N4 parameters

Build

• pyproject.toml: entry-point swap (linum_normalize_z_intensity → linum_correct_bias_field), scipy-stubs/networkx-stubs dev deps, lint exemptions for linumpy/gpu/* and N4 diagnostic scripts