Create a separate package that will expose advanced OCR features

[msluszniak]

After discussion, we decided to not support advanced OCR features like available her: https://github.com/software-mansion/react-native-executorch/tree/3f83e48a18bdb5c73ff5d65d364129b29e9a1578 directly in RNE, but rather expose only a set of simple OCR features and as a follow-up create a separate package on top of them. This will keep RNE cleaner and will enable us more in-depth approach.

[msluszniak]

Scoped #1322 down to detect + recognize, so what stays in RNE matches what the PaddleOCR reference SDK does. Removed 1615 lines; the code below is preserved at the tree linked above and its exporters stay on the converge-ocr-exporter branch of export-scripts, so the follow-up package can lift them as-is.

Moved out of RNE, to build on top later

Document helper models (documentModels.ts, 507 lines, one fused .pte per backend)

• Page orientation classification (PP-LCNet doc_ori), 4-way rotation with a confidence gate
• Dewarping (UVDoc) — the model emits a sampling grid, the client applies it via the native warpByGrid op
• Table structure recognition (SLANet_plus): an encode step plus an autoregressive decode loop over structure tokens, and the HTML cell fill that maps OCR lines into the recognized skeleton
• Registry: models.documentModels.PP_HELPERS, the SLANet structure vocab

Layout analysis

• PP-DocLayoutV3 (RT-DETR) composed as an object detector, per-region OCR on native crops, DocumentBlock assembly with reading-ordered blocks
• Containment-based region merging, because that head emits nested duplicates and IoU NMS cannot separate them
• Registry: models.layoutDetection.PP_DOCLAYOUT, DOC_LAYOUT_LABELS

Vertical text

• Stacked-glyph grouping, per-glyph column reading (composing a column into a horizontal strip makes the recognizer hallucinate words), char-level re-detection of merged stacks
• Per-run options: vertical, tallCropRatio, maxRedetections, verticalMinConfidence

Notes for whoever picks this up

• TextBoxExtractor lost its supportsCharLevel flag and the charLevel argument, but the native extractCraftTextBoxes op still takes charLevel, so per-glyph boxes remain available without a C++ change.

[msluszniak]

Review follow-up on #1322 removed two more things the comment above assumed would survive, so the note about charLevel no longer holds.

extractCraftTextBoxes no longer takes charLevel. The branch was unreachable once the per-glyph reader went away (makeCraftExtractBoxes never exposed the flag, so false was the only value the library could pass), so the subtract plus erode/dilate path and the angle = 0 override came out with it. Per-glyph boxes now need that C++ branch restored, not just a TS argument.

rotate, crop and warpByGrid are gone from image_ops.cpp and from the cv ops surface. They were added for the orientation and dewarp stages and had no call sites left. Dewarping needs warpByGrid re-added alongside the model, since the sampling grid the model emits is useless without it. rectifyQuad stays, it is what the recognizer path uses.

Both are recoverable from the tree linked above.