Improve interruption responsiveness

[GeorgeNgMsft]

Before these changes, pressing Escape mid-query felt sluggish or unresponsive:

• The UI would show "cancelled" eventually, but only after the LLM finished responding
• During a slow/long generation (or a 429 retry pause), you'd be stuck waiting even though you
  already cancelled
• Streaming chunks could keep rendering in the UI for a moment after cancel

After these changes, cancel feels immediate:

• Press Escape → the HTTP connection to the LLM drops instantly
• No more waiting for the full response to come back before "⚠ Cancelled" appears
• Streaming stops mid-token rather than finishing the current response
• If you cancel during a rate-limit retry, you're freed immediately rather than sleeping the full
  backoff