Add AMD GPU secondary restoration (AMF Video SR + Real-ESRGAN)

[lyx178238820-source]

Summary

Adds a secondary restoration backend for AMD GPUs, so --secondary-restoration
is no longer NVIDIA-only. Two engines are provided behind one option:

:: AMD's own Video SR (default) - AMF filter sr_amf, D3D11 on Windows
jasna --input in.mp4 --output out.mkv --secondary-restoration amd-upscale

:: Real-ESRGAN RRDBNet, in-process on the ROCm device
jasna --input in.mp4 --output out.mkv --secondary-restoration amd-upscale ^
      --amd-upscale-engine realesrgan --amd-upscale-scale 2

:: --amd-upscale-scale accepts 2 / 4 / 6 / 8; 6x and 8x act as supersampling
:: (the restored crop is blended back at the mosaic size)
jasna --input in.mp4 --output out.mkv --secondary-restoration amd-upscale --amd-upscale-scale 8

Motivation

--secondary-restoration offers unet-4x, tvai and rtx-super-res - all NVIDIA.
On an AMD GPU the stage simply could not be used (the factory raised for any
non-none value), so AMD users only got the primary BasicVSR++ pass.

Design

Both engines implement the existing SecondaryRestorer contract
(restore(batch) -> batch, 256x256 crop in, 256 * scale out, blended back by
the pipeline), so nothing outside session_factory needed to change.

amf-sr - AMD Video SR 1.0/1.1 via FFmpeg's AMF filter sr_amf:
format=nv12, hwupload=derive_device=amf, sr_amf=..., hwdownload. AMF is AMD's
media framework and initialises over D3D11 on Windows (not Vulkan/ROCm);
--amd-upscale-algorithm selects sr1-0/sr1-1 (ML) or the non-ML fallbacks
bicubic/bilinear/point, --amd-upscale-sharpness maps to the filter's
sharpness. One FFmpeg process per clip, with stdin closed as soon as the
frames are written: the AMF filters buffer the tail of the stream and only
release it when the input ends, so a long-lived pipe stalls one frame short of
every clip (closing stdin per clip flushes deterministically and subprocess
gives a hard timeout for free).

realesrgan - Real-ESRGAN RRDBNet in-process on the ROCm device
(prefers_cpu_input = False, tensors never leave VRAM - the same strategy as
the RTX Super Res backend). The architecture is inferred from the checkpoint
(num_block/num_feat/pixel_unshuffle), so x4plus, x2plus and the 6-block
anime weights all load; a network stronger than the requested factor is
area-downsampled. Weights are auto-detected in model_weights/ or given with
--amd-upscale-model-path.

Plumbing

• new CLI group --amd-upscale-{engine,scale,algorithm,sharpness,model-path,ffmpeg-path,timeout}
• SessionConfig fields + AmdUpscaleEngineName literal
• session_factory: AMD branch, AMD-only guard relaxed
  (amd-upscale requires an AMD device; other backends still require NVIDIA)
• GUI: engine/scale/algorithm/sharpness/timeout rows, single-choice entry, presets
• locales: en / zh / ja / ko / th
• docs: docs/zh/models.md, docs/zh/cli.md
• tests: tests/test_amd_upscale_secondary_restorer.py (FFmpeg command + filter
  chain, mocked), tests/test_realesrgan_restorer.py (CPU, tiny checkpoints),
  plus GUI-settings coverage
• benchmark record: benchmarks/2026-10-06_amd_secondary_upscale.{md,csv}

Testing

• pytest on this branch: 76 passed, 2 skipped (the new backend tests are pure
  CPU / mocked FFmpeg, so they run without a GPU)
• End-to-end on Windows 11 + RX 7900 XT (torch 2.9 ROCm, FFmpeg 8.1.2 with AMF),
  15 s 1080p clip, YOLO detection, 4 restored clips: both engines exit 0 and
  produce the expected output.

Benchmarks (RX 7900 XT, 256x256 crops)

engine	scale	output	throughput	notes
amf-sr	2x	512x512	~84 fps
amf-sr	4x	1024x1024	~55-60 fps	no extra install, uses AMD's own model
amf-sr	6x	1536x1536	~31 fps	still super-resolved, acts as supersampling
amf-sr	8x	2048x2048	~17-18 fps	ditto
realesrgan	4x	1024x1024	~16.5 fps	fp16, batch 2 (batch 4 hits a pathological MIOpen kernel: 0.24 fps)

sr_amf takes a free output size, so 6x/8x needed proof that the engine does not
silently fall back to plain scaling. Comparing sr_amf against algorithm=bicubic at
the same output size (one testsrc2 256x256 frame): PSNR 36.8 / 38.0 / 38.6 / 38.8 dB
at 2/4/6/8x, with 1.6-3.1x the Laplacian variance of bicubic, and 8x SR differs from
"4x SR + plain resize" by 47.1 dB - the engine keeps doing its own work at the higher
factors. Because the blend step downsamples the restored crop back to the mosaic size,
the higher factors behave as supersampling (cleaner edges, less aliasing) rather than
as more output resolution, at a proportional time cost.

Notes / limitations

• amf-sr needs an FFmpeg build with the AMF filters (sr_amf); the Windows
  Gyan "full" build has it. Error messages point this out.
• realesrgan needs a Real-ESRGAN checkpoint in model_weights/ (not bundled);
  the error text says exactly where to put it.
• Verified on RDNA3 (gfx110x). RDNA2 needs a ROCm build that ships gfx103x
  kernels (ROCm 7.14 / TheRock multi-arch wheels) - a wheel matter, not code.