Allow forked agents to take multiple turns (interactive mode)

Problem

Forked agents (spawned via the Agent tool without a subagent_type) are currently limited to a single turn. They execute autonomously, return a result, and the parent conversation receives a completion notification. That's it.

This means forks have to guess at every decision point. If the prompt was ambiguous or the fork encounters a choice that could go multiple ways, it picks one — and if it picks wrong, the entire fork's work is wasted.

Proposal

Add an opt-in setting that lets forked agents take multiple turns by asking the user questions mid-execution. Instead of running as a single autonomous pass, the fork could:

1. Start working on the task
2. Hit a decision point or ambiguity
3. Surface a question to the user (via AskUserQuestion or similar)
4. Receive the user's answer
5. Continue working with that context
6. Repeat as needed until done

The fork keeps its full context across turns — same agent, same conversation history. The parent still gets the final completion notification when the fork finishes.

Why this matters

Forks are great for offloading work, but the one-turn constraint forces a tradeoff: either front-load every possible decision into the prompt (verbose, error-prone), or accept that the fork will guess and potentially waste effort. Multi-turn forks would let agents iterate toward the right outcome the same way the main conversation does.