KeyError: 'encoder' when loading recognition model with transformers 4.56+

[MattiasHaughom]

Issue:
surya-ocr==0.12.1 crashes with KeyError: 'encoder' when using transformers>=4.56.0.

Error location:
surya/recognition/model/config.py, line 17:
encoder_config = kwargs.pop("encoder") # KeyError when encoder not in kwargs

Root cause:
Newer transformers versions changed how configs are initialized during from_pretrained(). The encoder and decoder keys are no
longer passed in kwargs.

Suggested fix:
encoder_config = kwargs.pop("encoder", None)
decoder_config = kwargs.pop("decoder", None)

if encoder_config is None or decoder_config is None:
if hasattr(self, 'encoder') and hasattr(self, 'decoder'):
return # Already set by parent class
raise ValueError("encoder and decoder configs required")

Current workaround:
Pin to transformers==4.36.2, but this conflicts with modern packages requiring transformers 4.41+.

Question: Does version 0.17.0 fix this issue?