feat: add common human-in-the-loop patterns

[shentongmartin]

feat: Add Common Human-in-the-Loop Patterns

This PR introduces a suite of common and powerful "human-in-the-loop" (HITL) patterns to the Eino examples. These patterns demonstrate how to build sophisticated agentic workflows that can pause for human intervention, enabling collaboration, oversight, and fine-grained control over agent execution.

Each pattern is implemented as a self-contained example within the adk/human-in-the-loop/ directory and includes detailed README.md files in both English and Chinese.

Core Contribution

This PR adds four distinct HITL patterns:

1. Approval: Simple, explicit approval before a critical tool call.
2. Review-and-Edit: In-place editing of tool call arguments before execution.
3. Feedback Loop: Iterative refinement of generated content based on human feedback.
4. Follow-up: Proactive clarification when a tool's output is insufficient.

These examples are built using a new set of reusable tool wrappers (approval_wrapper.go, review_edit_wrapper.go, etc.) that make it easy to integrate these HITL mechanisms into any agent.

---

1. Approval Pattern

• What it is: The most straightforward HITL pattern. An agent's action (specifically, a tool call) is paused, and a human user must explicitly approve (Y) or deny (N) it before it can proceed.
• How it works: It uses an InvokableApprovableTool wrapper that triggers a stateful interrupt before executing the wrapped tool. The workflow only resumes after receiving the user's confirmation.
• Use Case: Ideal for critical or irreversible operations, such as database modifications, sending emails, or confirming financial transactions.

2. Review-and-Edit Pattern

• What it is: A more advanced pattern that allows a human to act as a final checkpoint with editing capabilities. The user can review the arguments for a tool call and either approve them, disapprove the call, or edit them in-place before execution.
• How it works: It uses an InvokableReviewEditTool wrapper. On interrupt, the user can provide a new JSON string with corrected arguments, which are then used when the tool call resumes.
• Use Case: Perfect for scenarios where an agent might misinterpret details. For example, correcting a destination in a flight booking, adjusting a parameter in a complex API call, or refining a database query.

3. Feedback Loop Pattern

• What it is: An iterative refinement pattern where an agent generates content, and a human provides qualitative feedback for improvement. The loop continues until the human is satisfied.
• How it works: A multi-agent setup where a WriterAgent generates content and a ReviewerAgent presents it to the user. The user's feedback is passed back to the WriterAgent to generate a revised version.
• Use Case: Content creation workflows like drafting emails, writing articles, generating code, or creating marketing copy, where iterative human guidance is key to achieving the desired quality.

4. Follow-up Pattern

• What it is: A proactive pattern where an agent can recognize when a tool's output is insufficient to complete a task and can ask the user for a follow-up action or clarification.
• How it works: The agent uses a FollowUpTool that, when its own capabilities are insufficient, can interrupt and ask the user for the next step (e.g., "The search results are ambiguous, which one should I proceed with?").
• Use Case: Complex research or data retrieval tasks where the results may be ambiguous or incomplete, requiring human judgment to guide the next step.