[codex] Optimize rigid iterative processing

[tctco]

Summary

• skip discarded VoxelMorph work for rigid-only requests
• preserve iterative rigid affine semantics while removing unused temp writes
• speed up percentile clipping and bump DCCCcore to 4.2.3-alpha
• document CT benchmark and rejected low-resolution-cache experiment

Validation

• docker compose -f docker-compose.core.yml run --rm dccc-core /workspace/scripts/docker-build-core.sh
• docker compose -f docker-compose.core.yml run --rm dccc-core ./install/bin/DCCCcore --version
• docker compose -f docker-compose.core.yml run --rm dccc-core pytest tests/test_normalize_cli.py -q
• docker compose -f docker-compose.core.yml run --rm dccc-core pytest tests/test_acc_centiloid_centaurz_cli.py -q
• docker compose -f docker-compose.core.yml run --rm dccc-core pytest tests/test_acc_fill_states_cli.py -q

Release

• Tag pushed: v4.2.3-alpha

[tctco]

追加更新：已将 rigid alignment 和 VoxelMorph warping 拆成独立 lazy-loaded 组件。\n\n本次提交：be37642 Decouple rigid and VoxelMorph normalizers\n\n要点：\n- RigidVoxelMorphNormalizer 现在只作为 facade，按实际路径懒加载 RigidAlignmentNormalizer 或 VoxelMorphNormalizer。\n- 删除旧的组合 RegistrationPipeline，避免运行路径重新同时初始化 rigid/VoxelMorph。\n- rigid 路径不再加载 VoxelMorph；normalize --manual-fov 不再加载 rigid。\n\n本地验证：\n- Docker build 通过。\n- tests/test_normalize_cli.py: 4 passed。\n- tests/test_acc_centiloid_centaurz_cli.py: 30 passed。\n- CT rigid iterative 当前耗时 6.270s，峰值 RSS 2,009,480 KB；与旧输出比较 sform/qoffset/quaternion 最大差异均为 0。\n- 坏 rigid model 配置下，normalize --manual-fov 成功，而 rigid 正常失败，确认 skip-rigid 不初始化 rigid engine。