This Pull Request improves the robustness of the multithreading in the inference pipeline.

Changes:

• Introduces _safe_put and _safe_get helper methods to handle Queue operations safely.
  • Queue operations no longer use a timeout.
  • The producer and consumer respond to a shared _stop_event, ensuring proper shutdown even if one thread crashes or the user interrupts execution (e.g., Ctrl-C).
• Solves rare queue.Full errors that could occur during model cold start or if the consumer is temporarily slow (reported here, happened to me on rare occasions as well).
• Ensures that both leftover batches and the sentinel (None) are enqueued safely without risk of blocking.
• Maintains batch order and correct exception propagation between threads.

Impact:

• Cleaner and more predictable shutdown of preprocessing and inference threads.
• Queue-related deadlocks are prevented.
• Users should no longer encounter queue.Full errors during asynchronous inference, even with large models or slower hardware.