SOTA OCR on-device with Core ML and dots.ocr

A state-of-the-art OCR model, `dots.ocr`, can now run directly on Apple devices, delivering performance that surpasses Gemini 2.5 Pro without requiring an internet connection. This on-device processing takes advantage of Apple's Neural Engine, a custom AI accelerator that is dramatically more power-efficient than a device's CPU or GPU. However, converting complex models to run on the Neural Engine via the Core ML framework is a difficult task filled with potential errors. The successful conversion process involves simplifying the model's code, such as standardizing attention mechanisms, and then methodically debugging issues like data type mismatches and unsupported operations.