Add inworld router

[maxkahan]

This pull request introduces asynchronous backend persistence for Stream Chat message syncing in the StreamConversation class, improving latency for voice and LLM-driven pipelines by dispatching REST calls as background tasks. It also ensures test and shutdown reliability by allowing callers to await completion of all in-flight syncs. Additionally, the changes update tests to await these background tasks and improve error handling and documentation.

Asynchronous backend persistence and ordering:

• Refactored StreamConversation to dispatch Stream Chat syncs as fire-and-forget background tasks, serializing them with an asyncio.Lock to preserve message ordering and reduce latency on the critical path. Added a wait_for_pending_syncs() method to allow draining of in-flight tasks during tests and shutdown. [1] [2] [3] [4]
• Updated the agent shutdown path to await pending conversation syncs, ensuring no in-flight writes are dropped.
• Implemented a no-op wait_for_pending_syncs() in the base conversation class, with an override in StreamConversation for actual draining.

Test reliability improvements:

• Updated all relevant tests in test_stream_conversation.py and test_message_chunking.py to await wait_for_pending_syncs() before making assertions, ensuring test correctness with asynchronous persistence. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21]

Robustness and documentation:

• Improved error handling for Stream API exceptions during sync.
• Enhanced docstrings and comments for clarity on asynchronous behavior and usage. [1] [2]

Bug fixes and pipeline improvements:

• Prevented empty final LLM responses from being persisted, avoiding conversation history corruption and downstream provider errors.
• Ensured TTS pipeline always signals turn boundaries, preventing hangs in streaming scenarios.

Documentation update:

• Expanded inworld/README.md to better describe plugin capabilities.