Fix loading of legacy HuggingFace BERT checkpoints.

[drivanov]

Some legacy HuggingFace checkpoints such as prajjwal1/bert-tiny do not contain a config.json with the model_type field required by recent versions of Transformers.
As a result, AutoModelForSequenceClassification.from_pretrained() fails with:

  File "/workspace/examples/llm/glem.py", line 461, in <module>
    main(args)
  File "/workspace/examples/llm/glem.py", line 88, in main
    tag_dataset = TAGDataset(root, dataset, hf_model,
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch_geometric/datasets/tag_dataset.py", line 89, in __init__
    self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/transformers/models/auto/tokenization_auto.py", line 773, in from_pretrained
    return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/transformers/tokenization_utils_base.py", line 1721, in from_pretrained
    return cls._from_pretrained(
           ^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/transformers/tokenization_utils_base.py", line 1910, in _from_pretrained
    tokenizer = cls(*init_inputs, **init_kwargs)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/transformers/tokenization_utils_tokenizers.py", line 341, in __init__
    raise ValueError(
ValueError: Couldn't instantiate the backend tokenizer from one of: 
(1) a `tokenizers` library serialization file, 
(2) a slow tokenizer instance to convert or 
(3) an equivalent slow tokenizer class to instantiate and convert. 
You need to have sentencepiece or tiktoken installed to convert a slow tokenizer to a fast one.

This PR adds a small fallback for such models by directly using BertForSequenceClassification, while keeping the default AutoModelForSequenceClassification path for all other models.

[codecov]

Codecov Report

❌ Patch coverage is 0% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 84.18%. Comparing base (c211214) to head (990e373).
⚠️ Report is 185 commits behind head on master.

Files with missing lines	Patch %	Lines
torch_geometric/llm/models/glem.py	0.00%	6 Missing ⚠️

❌ Your patch check has failed because the patch coverage (0.00%) is below the target coverage (80.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #10631      +/-   ##
==========================================
- Coverage   86.11%   84.18%   -1.93%     
==========================================
  Files         496      510      +14     
  Lines       33655    36017    +2362     
==========================================
+ Hits        28981    30321    +1340     
- Misses       4674     5696    +1022     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[drivanov]

@akihironitta: I see exactly the same issues with the code coverage check as in PR#10623. Otherwise, in my opinion, this PR is ready to merge.