Capture partial messages and usage when tasks are cancelled

[dedeswim]

Currently when a task is cancelled (Ctrl+C), we record the cancellation but cannot save the partial messages and token usage accumulated up to that point.

The agent's run() method uses a simple iteration pattern that doesn't expose intermediate state:

async for state in self.iter(...):
    result_state = state
return result_state.run_ctx

To capture partial state, we would need to either:

1. Store the last yielded state in an instance variable accessible on cancellation
2. Implement streaming with real-time state callbacks

This would be useful for debugging cancelled runs and understanding how far execution progressed before interruption.

[cneuralnetwork]

Implemented this @dedeswim !