refactor(ai): Adopt whilesmartgo/agents for the assistant core

[nfebe]

The assistant's engine, tool registry, harness loop, and tool authorization are implemented inline across internal/ai and internal/api, coupled to gin and FlatRun's auth. github.com/whilesmartgo/agents now provides that core as a provider-agnostic, framework-free library. Depend on it and drop the duplicated core, keeping only what is FlatRun-specific.

Checklist

• Add github.com/whilesmartgo/agents (pinned) as a dependency.
• Replace the internal/ai provider abstraction (Provider, Message, ToolCall, Request, Response) and the OpenAI-compatible client with the library's Engine and engine/openai.
• Express the built-in tools (list_deployments, restart_container, run_host_command, and the rest) as agents.Tool values in a Registry, keeping their existing FlatRun implementations.
• Drive the session tool loop through agents.Runner, supplying Authorize from the current actor and deployment-write gate and Approve from the awaiting-approval flow.
• Keep the FlatRun-specific parts: the gin handlers, the flat-file session store and its per-user scoping, suggested-action parsing, redaction, and the display/hidden transcript projection.
• Confirm the assistant behaves identically (auto-run, approval, tool errors, step budget) against the existing tests.

Related

• Library: https://github.com/whilesmartgo/agents