fix dimensions of image generated by parser.py

[montvid]

After image ocr with --no_fitz_preprocess a jsonl file is generated with wrong image dimensions to the original because smart_resize is used in the parser.py script. As per https://arxiv.org/abs/2307.06304 NaViT can ingest any dimension image so no need for smart_resize.

[ygfrancois]

Smart resize here is to show the real input size to model, the model server will online do the smart resize to keep the input size divisible by vision patch size.

[montvid]

Thanks for clarification! Now I understand that NaViT-style encoders accept arbitrary aspect ratios/resolutions, but the deployed implementation still enforces patch-size divisibility to form valid patch tokens efficiently.