Support automatic "back" handoffs to orchestrating Agents

[niv-hertz]

Currently, handoffs are a one-way ticket. Once an Agent handed off control to another Agent, it will never regain control. The only way to return control to the previous agent, is to have some circular hand off mechanism (which would require a circular reference between the two agents, which is far from ideal). There are scenarios, where you don't even need to call an LLM to determine that you want the main agent to regain control, assume this architecture:

A simple user question may require more than one agent to run, Ideally, you'll want Agent A. to do the work, go back to the main agent, and then call Agent B. The current solution to this, is to use Agent A and Agent B as tools, but that's not ideal for a couple of reasons (Streaming is not easily supported, context is limited, cannot limit the max turns).

We can generate a similar architecture with Handoffs, by implementing an optional back_handoff parameter, that will result in an automatic Handoff back to the main agent, whenever one of the sub-agents complete their task.

[niv-hertz]

@rm-openai

I'll be happy to implement this, I thought about adding an optional parameter to the Handoff definition, that controls the back_handoff logic.

The implementation itself will be in the Runner - once an handoff with back_handoff definition is invoked, it will retain a reference to the previous agent (as a stack). Once a NextStepFinalOutput output item, is returned by _run_single_turn, check the stack to see if any handoff with back_handoff was performed, if so, hand it back to that agent instead of ending the loop.