Allow subagents to finish with final response output

Currently all subagents today get a `complete_task` tool. For subagents that do not have or want a more complex return schema, an absence of schema could be used to omit this tool. In this case, the subagent is considered finished when no more tool calls are made.

Considerations:
- The model could return with less helpful or no response text at all. In this case, there is a less useful handoff.
- Ensure this does not reduce quality for any built-in subagents
- Behavior may vary by model

Ideally, we can support both completions via the tool and without depending on the schema used.