fix: recover orphaned assistant messages after restart

[Haohao-end]

Issue for this PR

Closes #19023

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

This fixes a case where a session can stay stuck in a thinking state after the server restarts or a stream is interrupted.

Before this change, unfinished tool parts were only cleaned up at the end of the normal processor flow. If the process exited before that cleanup ran, the assistant message could remain incomplete and its tool parts could stay in pending or running.

This change extracts that cleanup into a reusable session helper and reuses it in two places:

• the existing processor shutdown path
• instance bootstrap, so interrupted assistant messages are recovered when the instance starts

During recovery, unfinished tool parts are marked as error and the assistant message is completed. This reuses the existing message/part update flow so clients receive the normal update events instead of staying stuck.

How did you verify your code works?

I verified the change by tracing the existing session update flow and confirming that the recovery path uses the same updatePart() and updateMessage() calls as the normal cleanup path.

I was not able to run the full local typecheck in my current environment because the typecheck script depends on tsgo, which was not available (tsgo: command not found).

Screenshots / recordings

Not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍

[Haohao-end]

Rebased on latest dev and all checks are now passing. Happy to make follow-up changes if needed.