fix: stop RAM growth during full pipeline runs

[tqbdev]

Problem

Running the pipeline over all pages (detect → OCR → translate → inpaint → render) made RAM climb monotonically page after page and never release it, building to large consumption by the end of a batch. Reproduced on macOS with the LaMa inpainter on Metal.

Root causes

Three independent leaks (a fix for any one alone barely moved the needle):

1. Frontend — object URLs never revoked. useBlobImage (ui/hooks/useBlobData.ts) created URL.createObjectURL(blob) but never revoked it, so every sprite/inpaint/render blob pinned a full decoded image in the webview for the whole run. Engine-independent, so backend changes didn't help it.
2. Backend — no Objective-C autorelease pool on Metal. candle's Metal ops and the MPSGraph FFT (lama/fft/metal.rs) allocate autoreleased objects (command buffers, MPS intermediates); a leaked command buffer also pins every GPU buffer it referenced. The pipeline runs on tokio worker threads that have no autorelease pool draining, so these accumulate for the entire run. There was no autoreleasepool anywhere in the tree.
3. Backend — unbounded ML caches. The LaMa FFT-plan caches (keyed by crop shape, lama/fft/{metal,cuda}.rs) and the flux2 Qwen prompt cache (flux2_klein/qwen.rs) lived on whole-run engine instances and were never evicted.

Fixes

• ui/lib/queryClient.ts: revoke a blobImage query's object URL when it's evicted from the query cache; ui/hooks/useBlobData.ts: drop blob gcTime 10min → 1min so inactive pages release promptly.
• lama/mod.rs + lama/fft/metal.rs: wrap the LaMa per-crop forward and the FFT calls in objc2::rc::autoreleasepool (no-op off Metal via a small cfg-split helper).
• lama/fft/{metal,cuda}.rs and flux2_klein/qwen.rs: bound the caches with lru::LruCache.

Verification

• cargo check -p koharu-ml passes with and without the metal feature.
• cargo test -p koharu-ml --features metal --lib lama:: passes.
• UI lint clean (no new warnings).
• Manually confirmed on macOS/Metal: RAM now plateaus during "process all pages" instead of climbing to the end.

Notes

• If growth ever shows up specifically during Detect/OCR (not Inpaint), it's the same candle-Metal autorelease issue in those engines and can be fixed with the same wrap.

🤖 Generated with Claude Code

[github-actions]

Thanks for your first PR to Koharu.

Please review our contribution guide before review:
https://koharu.rs/contribute/introduction/

In the PR description, include:

• what changed
• any user-visible behavior differences
• how you verified the change

If AI helped produce the patch, a human still needs to review and understand it before submission.