on-device upscaling mvp: ncnn spike + phase 1

[sleepy-snowflake]

Implements Phase 0 (native spike) + Phase 1 (on-device MVP) per docs/phases.md.

Spike (Phase 0)

• Vendored ncnn 20260526 android-vulkan prebuilt (arm64-v8a + armeabi-v7a) under app/libs/ncnn/ with VERSION.txt provenance
• CMake + externalNativeBuild; JNI bridge (upscale_jni.cpp / NcnnBridge.kt): init/loadModel/upscaleTile/release, error-code semantics only
• Bundled realesr-animevideov3 x2+x4 ncnn models (+meta.json, BSD-3-Clause)
• Debug-only "Engine self-check" smoke screen: one 256px tile before/after with GPU/CPU ms timings

MVP (Phase 1)

• Engine toggle (Cloud | On-device) + model dropdown on UploadScreen
• Oversize dialog (>1024px long edge): shrink-and-upscale (recommended), one-shot cloud w/ connectivity gate, full-res gated by preflight heap math (greyed + reason when impossible)
• Full local pipeline: decode -> EXIF -> sRGB -> alpha flatten -> replicate-filled 256px tiles -> NCNN inference -> feathered blend -> JPEG encode -> existing Downloads save path
• Honest progress via existing Processing state ("On-device · tile n/N · GPU|CPU"), cancel between tiles
• 4096px output cap enforced in DownscaleFit; largeHeap enabled; sticky GPU->CPU fallback
• Unit tests: tiler coverage/seams/blending + preflight math (CI now runs testDebugUnitTest)

Notes

• UpscaleState untouched; ApiService.kt untouched except cancellation guards in UpscaleViewModel cloud branch
• docs/phases.md checkboxes updated; architecture.md adjusted (replicate-fill tiles instead of pad+crop, actual blob names data/output)
• ARM-only ncnn: physical device testing required (emulators won't run it)

Device QA checklist

• Camera photo -> correct-side-up sharp x4 output
• Low-RAM device survives big photos
• CPU path exercised (airplane mode)
• Cloud flow regression check