Fix: 修复首张文本场景图重复OCR问题，更新关于页技术栈徽章

Revert: 回滚ONNX(HF Spaces上反而更慢), 移除可视化文字叠加

Perf: PaddleOCR ONNX Runtime 加速，文本场景提速 ~20%

Perf: PP-OCRv3 + fast scene classify for CPU speedup

Fix: clean Qdrant cache when history empty, remove retest button

Fix: demo skip classify_scene, fix unpack error, fix cover path, update track name

Add cover bg, scene mode selector, single image upload, demo hides upload, about page format

Glassmorphism UI, about page, real demo images, skeleton loading, parallax

Add demo scene button, thread config, user expansion

Fix: remove rect boxes from vis, pass per-image crop_offset, only draw lines

Fix: patch coord mapping with CenterCrop offset, PIL for Chinese text in text vis

Fix visualization: build before delete, add text scene vis, click-to-show

Fix: pass patch_match_info through evaluator to enable visualization

Add: DINOv2 threshold 0.5, patch match visualization, history image management

Fix paddlepaddle version to 2.6.2