fix(redact): detect 512² input so rfdetr_v11 actually runs (was 384 fallback)

v11 ships a 512×512 ONNX, but detect_input_size()'s allowlist was still
[320, 384]. Detection returned None → load() fell back to .unwrap_or(384),
so every frame fed a (1,3,384,384) tensor to the 512 model:
`ort run: Got invalid dimensions for input` on a ~2s backoff loop. Net
effect on v2.4.305: the image PII model loads but fails every inference —
image redaction silently does nothing.

Add 512 to the allowlist (+ comment to keep it in sync on the next model
bump). Verified against the real on-disk rfdetr_v11.onnx: load + detect()
now run clean (previously errored on dimensions).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>