convert : fix bytes_to_unicode import for transformers >= 5.15

[SolshineCode]

Overview

convert_hf_to_gguf.py fails with an ImportError on transformers >= 5.15 because bytes_to_unicode moved from transformers.models.gpt2.tokenization_gpt2 to transformers.convert_slow_tokenizer. Any model that goes through QwenModel.token_bytes_to_string hits it. This adds a try/except: old path first, new location as fallback, so both transformers ranges work.

Additional information

Hit while testing #26185 (the Kimi K3 tokenizer goes through this path). Verified on transformers 5.15.0.dev0: conversion completes and the resulting GGUF matched my HF reference outputs.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - the fix was found during AI-assisted testing (Claude, disclosed in my model: add Kimi-K3 text model #26185 test report); I reviewed and tested the change and take responsibility for it.

[SolshineCode]

Rewrote the description and commit message per the guidelines.

[CISC]

The rationale is flawed, it wasn't moved, it was simply removed in huggingface/transformers#40936 as the function was already copied to this location in huggingface/transformers#30334 (4.54.1), meaning it is safe for us to use it from there.

Please make the same fix in chatglm.py.

[SolshineCode]

Thanks, you're right that it wasn't moved. It was removed in #40936 and had already been copied to convert_slow_tokenizer in #30334 (4.54.1). I've now corrected the commit message. I've also switched to importing directly from convert_slow_tokenizer, and applied the same fix to chatglm.py.