AI: Real-ESRGAN texture upscaling

[fernandotonon]

Parent epic: #397

Goal

Add 2×/4× texture super-resolution using Real-ESRGAN via the ncnn-Vulkan port (or ONNX as fallback). Useful for upscaling low-resolution imported textures and sd.cpp outputs.

Scope

• Pick the backend:
  • Preferred: realesrgan-ncnn-vulkan (native C++, cross-platform GPU).
  • Fallback: ONNX Runtime if Vulkan availability is a concern on macOS arm64.
• CLI: qtmesh material --texture low.png --upscale 4 -o high.png.
• GUI: Material Editor right-click on a texture slot → "Upscale 2×/4×".
• MCP: upscale_texture tool.
• Cache + skip re-upscale of already-upscaled images.

Acceptance Criteria

• 4× upscale of a 256×256 albedo produces a sharp 1024×1024 image.
• Works on macOS arm64, Windows MinGW, and Linux CI.
• Falls back gracefully when the backend is unavailable (clear error, no crash).
• Sentry breadcrumb ai.assist.upscale recorded.

Effort

~3–4 days.

[fernandotonon]

Shipped in #749 (squash-merged to master).

Real-ESRGAN texture super-resolution (2×/4×), ONNX-backed, reusing the #404 ONNX Runtime infra:

• TextureUpscaler (Ogre-free, unit-tested core): scale-aware overlapping-tile upscale, compositing in OUTPUT space with a feathered seam blend; scale factor detected from the model's output/input ratio at runtime (validates uniform-integer scale on both axes); alpha preservation; output-allocation guard (256 Mpx cap, bad_alloc handling).
• Model: Real-ESRGAN x4plus / x2plus (BSD-3-Clause, xinntao), exported to ONNX via the offline-only scripts/export-realesrgan-onnx.py (with SHA-256 verification before deserialization), hosted at fernandotonon/QtMeshEditor-models, auto-downloaded on first use.
• Surfaces: CLI qtmesh material --texture <low> --upscale {2|4} [-o <high>], MCP upscale_texture, and Material Editor "Upscale 2×/4×" buttons.
• UX: download-vs-upscale status, per-tile progress, and a Cancel button (worker-threaded; cancel via shared atomic checked in the progress callback).
• Performance: ONNX intra-op threading set to hardware_concurrency-1 — a 256→1024 4× dropped from ~2 min to ~7.5 s (~7 cores) on an M-series laptop. CoreML EP on macOS, CPU EP elsewhere.

All acceptance criteria met; verified end-to-end (256→1024 4×, 128→256 2×) with the model auto-downloaded. Closing.