Evaluate standalone agents

[clareliguori]

I'd like to be able to build a standalone agent using the tau2 domain tools, then evaluate the standalone agent using tau2.

A few changes I can think of that would be involved:

1. Separate the tools out into something like an MCP server.
2. Remote agent invocation: Similar to Improving A2A Agent Integration for tau2-bench #111, it could be invoked over A2A.
3. Offline evaluation: Ability to convert the remote agent's state/messages into a format that tau2 can evaluate after a run.