Fix Conversational AI audio chunk handling and WebSocket error recovery

[18-RAJAT]

• Add audio chunk queue/buffer to prevent loss during transient WebSocket failures
• Improve error handling to distinguish between fatal and recoverable errors
• Add connection state checking before sending audio chunks
• Add proper exception handling for WebSocket errors (including 1011 internal error)
• Use dedicated thread/task for sending audio chunks to prevent blocking
• Handle ConnectionClosedOK, ConnectionClosedError, and ConnectionClosed exceptions gracefully
• Prevent session termination on transient errors to improve reliability

[kraenhansen]

First off - thanks a lot for your interest in the project!

This PR addresses a bit too many issues at once for us to reliably review and merge - especially since this is your first PR to our projects. Could I get you to separate this out into:

• an initial PR with the refactoring (no new features or actual bug-fixes),
• then PRs for each bug that you're fixing
• and finally a PR for the audio chunck queue / buffering.

This will make it easier for us to review and potentially partially revert if this (against expectations) introduce new bugs.

[kraenhansen]

Hey! We're doing a spring cleaning of open PRs. It looks like there are some outstanding requested changes or questions on this PR. Is it still something you'd like to get merged? If so, could you update it when you get a chance? We're marking this as stale for now and will close it in a week if there's no activity.

[kraenhansen]

Also, your branch has merge conflicts and will need a rebase before it can be merged.