perf(layout): Process multi-frame images one frame at a time - #520

Large multi-frame images currently retain every decoded RGB frame before layout inference starts. Process each frame through inference immediately and close its converted buffer after the page completes. Page order, dimensions, format metadata and pixel contents are preserved.

Validation:

• Full local inference suite: 269 passed, 1 skipped; two added cleanup/streaming regressions also pass inside the dedicated SND.
• Independent review checked ten baseline/candidate cases across TIFF color modes, GIF/APNG, varying frame dimensions and error paths.
• Dedicated nick-oom-perf-oct01 SND, same core-product API image digest, separate Linux pods limited to two CPU cores: three matched trials on 160 compressed 1024×1024 TIFF frames produce identical output hashes. Median decoding time is 0.392 → 0.176 seconds; peak process RSS is 869 → 232 MiB. This isolates decoding using an empty fixed layout; it is not an end-to-end model latency claim.
• At a 768 MiB pod limit, the baseline is OOMKilled (exit 137), while the candidate completes the same fixture.
• Real YOLOX inference on a three-frame text TIFF also produced matching baseline/candidate output in the SND.

A corrupt later frame is now discovered after inference on earlier valid frames; the tested corrupt input raises the same exception and neither implementation returns a partial document. Decoded page layouts still accumulate in the returned document, so memory remains proportional to output size.

Release metadata targets 1.6.15. Sibling PRs take #518 → 1.6.14, #521 → 1.6.16 and #522 → 1.6.17; merging out of that order requires renumbering. Consumers require the resulting inference-library release to receive the fix.