docs: add 'Run on a single 24 GB GPU' section (4/8-bit, torch>=2.6 + gcc reqs, verified results)

Add Lapa Ukrainian HW-OCR LoRA adapter v1

initial commit