This includes the tricky case, where multiple tool calls are launched
and *some* of them return results and some need hooks.

Approach:
 * Change hook aborts to raise an Exception instead of cancelling the task
 * Callers should handle the exception, probably by creating a
   `pending_tool_result` that has `is_hook_abort=True`
 * `Agent.run`'s replay handling gets extended to detect this case
   and merge the tool results into the assistant message for replay.
   (Maybe this new function, `_process_interrupted_hooks` should just
   be user facing?)