Release 1.4.0

Fix for long inputs crashing inference, plus documentation and
dependency updates.

• Fixed a crash on long descriptions: some fine-tuned repos (notably the
  Chinese MacBERT model) ship no model_max_length in their
  tokenizer_config.json, so transformers reports a huge sentinel value
  and truncation=True never actually truncated — inputs over 512
  tokens overflowed the model's position-embedding table at inference
  time. Both classifiers now clamp the tokenizer's limit to the model's
  max_position_embeddings (minus the two offset slots RoBERTa-style
  models reserve) right after loading, via a shared
  clamp_tokenizer_max_length helper with its own test suite.
• Documentation: the README now shows HF_HUB_OFFLINE=1 as an example
  of starting the server without pulling model updates from the
  Hugging Face Hub.
• Updated dependencies.