convert : register Qwen 3.5 ForCausalLM for text only

[CISC]

Support text only versions of Qwen 3.5.

Fixes #20116
Fixes #20102

[danbev]

This does fix the reported issue but the conversion fails with the following error for me:

Traceback (most recent call last):
  File "/home/danbev/work/ai/llama.cpp/examples/model-conversion/../../convert_hf_to_gguf.py", line 12130, in <module>
    main()
  File "/home/danbev/work/ai/llama.cpp/examples/model-conversion/../../convert_hf_to_gguf.py", line 12124, in main
    model_instance.write()
  File "/home/danbev/work/ai/llama.cpp/examples/model-conversion/../../convert_hf_to_gguf.py", line 716, in write
    self.prepare_metadata(vocab_only=False)
  File "/home/danbev/work/ai/llama.cpp/examples/model-conversion/../../convert_hf_to_gguf.py", line 857, in prepare_metadata
    self.set_vocab()
  File "/home/danbev/work/ai/llama.cpp/examples/model-conversion/../../convert_hf_to_gguf.py", line 829, in set_vocab
    self._set_vocab_gpt2()
  File "/home/danbev/work/ai/llama.cpp/examples/model-conversion/../../convert_hf_to_gguf.py", line 1336, in _set_vocab_gpt2
    tokens, toktypes, tokpre = self.get_vocab_base()
                               ^^^^^^^^^^^^^^^^^^^^^
  File "/home/danbev/work/ai/llama.cpp/examples/model-conversion/../../convert_hf_to_gguf.py", line 1005, in get_vocab_base
    tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/danbev/work/ai/llama.cpp/venv/lib/python3.11/site-packages/transformers/models/auto/tokenization_auto.py", line 1153, in from_pretrained
    raise ValueError(
ValueError: Tokenizer class TokenizersBackend does not exist or is not currently imported.
make: *** [Makefile:41: causal-convert-model] Error 1

But this looks like a custom class and with the following the conversion worked:

diff --git a/tokenizer_config.json b/tokenizer_config.json
index 6be6ce1..a5adeb7 100644
--- a/tokenizer_config.json
+++ b/tokenizer_config.json
@@ -23,7 +23,7 @@
   "pad_token": "<|endoftext|>",
   "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
   "split_special_tokens": false,
-  "tokenizer_class": "TokenizersBackend",
+  "tokenizer_class": "Qwen2TokenizerFast",
   "unk_token": null,
   "video_token": "<|video_pad|>",
   "vision_bos_token": "<|vision_start|>",

[CISC]

This does fix the reported issue but the conversion fails with the following error for me:

Thanks for testing! :)

I think this is a new class in transformers 5...

Edit: Not sure about the validity of using it like that, but seen it before, also a bit strange with the pretokenize_regex in tokenizer_config.json, likely that this will give a different chkhsh, but don't think I'll bother with that in this PR.