Implement PP-OCRv5 using official model names

Enable PP-OCRv5 support with fallback

textsense-ocr: set PPOCR_HOME to writable /tmp and create it to avoid permission errors

textsense-ocr: add libgl1 and libgomp1 to fix OpenCV/Paddle runtime deps

Switch OCR backend to PaddleOCR\n\n- Replace MiniCPM inference with PaddleOCR (CPU)\n- Simplify requirements to Paddle and PaddleOCR only\n- Use slim Python image; drop CUDA dependencies\n- Keep API compatible: /extract returns text lines\n

Replace Qwen2-VL with MiniCPM-V-4 for OCR functionality

Fix Qwen2-VL model loading - use correct model class

Fix dots.ocr model error by switching to Qwen2-VL-2B-Instruct

Switch to PyTorch devel image and set CUDA env vars for flash_attn compilation

Add flash_attn back and install build tools for dots.ocr model compatibility

Update transformers to >=4.55.0 for dots.ocr compatibility

Fix build: remove flash-attn dependency and add git to Dockerfile for dots.ocr compatibility

Fix dependencies: pin exact versions and update PyTorch base image for dots.ocr compatibility

Add missing dependencies for dots.ocr VLM model (transformers, accelerate, torch, etc.)

Replace TrOCR with rednote-hilab/dots.ocr; add VLM pipeline and dependencies

OCR: switch to TrOCR printed model and better decoding (beam search)

Improve error handling for network issues and connection failures

Add TrOCR OCR functionality with FastAPI - Docker container with TrOCR model for image-to-text extraction

initial commit