refactor(cuda): bring the CUDA kernels and their headers to the lint and HISS standard (ADR-1142)

[lusoris]

Summary

Brings the cuda clang-tidy lane to zero findings (581 before, 0 after; measured in the dev container, clang-tidy 22.1.8, scripts/dev/tidy-lane.sh --write cuda). The 62 files are the CUDA kernels (.cu, .cuh), their CUDA-only headers, headers shared with the C hosts, two CUDA tests and the shared picture.h DEVICE_CODE branch. scripts/ci/tidy-baseline-cuda.json is written by the tool, not by hand. Uncited NOLINT: 0.

• Device pointers are rebuilt from a CUdeviceptr with VMAF_CUDA_DPTR(T, address) (core/src/cuda/cuda_device_ptr.cuh).
• Helpers and device structs live in anonymous namespaces; __global__ kernels stay extern "C".
• No designated initializers in device code (nvcc MSVC host frontend, preflight.sh --stage msvcism passes): aggregates are filled field by field.
• Long functions are split (vif_statistic_calculation into vif_sigmas, vif_gain, vif_accumulate_log, vif_statistic_pixel; SpEED, CAMBI, ADM, SSIM, MS-SSIM, PSNR-HVS helpers).
• Headers shared with C carry the ADR-1138 / ADR-1470 NOLINTBEGIN brackets, as the cpu lane did.
• Source-text contract tests follow the new spelling; their assertions are unchanged.

Behaviour check

• Twin scores at --precision max, 20 extractors on --backend cuda: Netflix 576x324 pair (48 frames), both 1080p checkerboard pairs, a 10-bit 480x270 pair; 0 of 3127 values differ from master, pooled metrics equal.
• All 264 libvmaf tests with a cuda / gpu / vif / adm / ssim / psnr / motion / cambi / ciede / moment / speed name pass on the RTX 4090 under the lock, including every test_cuda_*_parity and *_large.
• sm_89 SASS before/after: identical for 15 of 21 kernel TUs; the rest differ by commutative operand order, register names, unsigned compares on non-negative indices, or (filter1d.cu) the split of the VIF statistic; fp64 and conversion instruction counts of filter1d are identical, LDG.E.CONSTANT counts equal everywhere.
• 4K wall times (bbb 200 frames, 18 features, 2 reps) within noise; the machine was loaded by other lanes.
• Netflix golden gate: not run; no CPU extractor code changed (comments in integer_vif.h, ssimulacra2_eotf_lut.h, NOLINT brackets only).

NOLINTs added

• core/src/cuda/cuda_device_ptr.cuh: performance-no-int-to-ptr on the VMAF_CUDA_DPTR macro (ADR-1142). A CUdeviceptr is an integer by the C ABI of the kernel argument structs; an inline function in place of the macro turns 134 of speed_score.cu's 378 LDG.E.CONSTANT loads into plain loads (measured, sm_89, nvcc 13.4).
• ADR-1138 / ADR-1470 brackets (modernize-use-using, performance-enum-size) in C headers: ssimulacra2_cuda.h, integer_adm_cuda.h, integer_psnr_hvs_cuda.h, integer_cambi_cuda.h, speed_cuda_params.h, integer_vif_cuda.h, integer_vif.h, float_vif_gpu_common.h, float_vif_device.h, float_adm_gpu_common.h, float_adm_device.h, ciede_device.h, libvmaf_cuda.h.
• ssimulacra2_eotf_lut.h (and its generator): modernize-use-std-numbers around the table; entries are EOTF samples, not constants.
• integer_vif.h: cert-dcl03-c,misc-static-assert on a runtime assert() (ADR-1142 forbids weakening it).

Checklist (ADR-0108)

• Research digest: no digest needed: lint refactor, no behaviour change
• Decision matrix: no alternatives: only-one-way fix (macro vs function measured above)
• AGENTS.md invariant note: core/src/cuda/AGENTS.md section "Device pointers and device-code lint"
• Reproducer / smoke-test command: scripts/dev/tidy-lane.sh cuda exits 0; meson test -C build --suite fast+gpu on a CUDA build
• CHANGELOG fragment: changelog.d/changed/tidy-zero-cuda-lane.md
• Rebase notes: docs/rebase-notes.md entry added
• docs/state.md: no state delta: lint refactor, no bug found or closed
• Netflix golden data: no assertAlmostEqual touched