Fix tokenizer saving for newer transformers versions

[alperctnkaya]

Summary

Fix compatibility with newer transformers versions where Trainer.tokenizer is replaced by processing_class.

The current implementation crashes during checkpoint saving with:

AttributeError: 'EncoderOnlyEmbedderM3Trainer' object has no attribute 'tokenizer'

This PR:

• keeps backward compatibility with older transformers versions
• supports newer versions using processing_class
• prevents checkpoint save failures during training

Tested

• transformers==5.9.0
• Python 3.12
• BGE-M3 fine-tuning

[hanhainebula]

Thanks for your contribution! I’ve provided a more comprehensive fix for this issue in #1592, so I’ll be closing this PR. I appreciate your effort and help with improving Transformers v5 compatibility.