fix(preview): bound GPU, thumbnail, and prefetch memory

[hsnilsson]

AI-assisted comment (model: GPT-5).

Summary

• Decode difficult thumbnails and LinearRaw previews with bounded background work so the UI stays responsive and duplicate foreground/background decoding is avoided.
• Tile GPU preview work, account for auxiliary cache buffers, and release render resources deterministically.
• Keep neighbor prefetch automatic, but run it only when preview-cache capacity and available system memory provide enough headroom. Serialize background decoding and cancel stale work during navigation.
• Bound IR downsampling, preserve IR defects across streamed RGBI segment boundaries, and keep half-frame native dimensions correct when previews use reduced RAW decodes.
• Update the user guide, pipeline documentation, and regression coverage.

Verification

• Ruff formatting and lint passed.
• Ty type checking passed.
• 577 focused controller, loader, cache, GPU, IR, half-frame, and thumbnail tests passed.
• Full suite with UTF-8 mode: 5,730 passed, 26 skipped, and 14 deselected.
• One unchanged upstream Windows test remains failing: tests/test_crosstalk_profiles.py::test_saving_again_keeps_the_profile_in_its_own_file. The TOML cache can reuse stale data when two same-length writes receive the same file modification timestamp. This branch does not change that test or the crosstalk/cache implementation.
• Large segmented LinearRaw DNG files were exercised through the bounded foreground and prefetch paths. Cache/RAM rejection happened before decoding, prefetch work stayed serialized, cancellation completed cleanly, and no OOM, GPU-validation, device-lost, or Qt thread errors were found.
• Explicit full-resolution HQ decoding remains limited by LibRaw's native full-frame CPU allocation and cannot be cancelled inside the native call.

[hsnilsson]

I'll split this into 2-3 smaller PRs. Don't waste time on reviewing just yet.
This is on the path to resolving my major blockers when handling large pixelshift images. I would normally speak in person with some other team member before pushing something like this ;)

[hsnilsson]

The other PRs are now pushed and should be merged in order #1107 -> #1119 -> #1121 -> #1122