Summary

• change the library dependency from the transformers 5.0.0 release candidate to stable transformers>=5.0.0
• update mlx-lm to 0.31.1, the latest non-yanked published release compatible with stable transformers 5
• require huggingface_hub>=1.0, which is a requirement of transformers >=5.0.0 anyway
• align the tokenizer tests with the current AutoTokenizer implementation and make the Llama shape assertions depend on the constructed input length

Verification

• uv lock
• uv run --extra dev --extra tts pytest mlx_audio/tests/test_optional_deps.py mlx_audio/tests/test_lazy_imports.py mlx_audio/tts/tests/test_models.py -q