Save each result (image + braille + english) to braille-reader-results dataset

Fix: use @spaces.GPU for ZeroGPU - load models on CPU, move to GPU per-request

Initial app: full braille OCR pipeline (YOLOv8 + ByT5)

initial commit