Add two task processing modes: queue and immediate interrupt

[Gustave-241021]

Specific implementation and plan : #1604

• Basic functionality: supports users with more flexible task handling options

Bonus:

Task Queue Behavior Change

This change addresses an issue where queued messages were visible to the currently executing task, potentially disrupting its execution. The new deferred queue ensures that queued tasks do not interfere with the active task, maintaining clear task boundaries and predictable behavior.

When a task is queued with DeferredQueue=true, it will only be processed after the current task completes. This approach:

• Prevents task interference: Queued messages are not injected into the active task's context, ensuring the current task completes as intended
• Enables corrections: Follow-up messages and corrections are still processed, but only after the current task finishes, allowing users to refine their prompts based on the completed task's results
• Improves predictability: Users can queue multiple tasks with confidence that each will execute independently, making the system's behavior more predictable and easier to debug

This design aligns with how major AI agent products handle message queuing:

• ChatGPT/Claude: Messages are processed sequentially; new messages wait for the current response to complete
• GitHub Copilot Chat/Cursor: Tasks execute independently with clear boundaries
• AutoGPT/AgentGPT: Tasks run serially without interruption

Future expansion can be referenced: #1412