Feature Request: Allow selection of smaller Hugging Face models in docstrange local mode

[sadhikariSteep]

When using docstrange in GPU local processing mode (via the DocumentExtractor), it appears to default to a specific large model nanonets/Nanonets-OCR-s.

For users with limited compute resources (e.g., less VRAM), running these larger 3B/4B parameter models locally can be slow.

For example, Nanonets provides models like nanonets/Nanonets-OCR2-1.5B-exp, which is significantly smaller and would be more accessible for local use on a wider range of hardware.

[linear]

PLG-207