Add support for Evals

[scottpersinger]

We should spec out an easy Eval system for testing agents:

• Write tests using simple Yaml
• Input prompt and expected output
• Indicated expected tool calls

Like:

name: Weather Agent Eval
evals:
  - it: should not know the time
     input: What time is it?
     eval_response: I don't have access to time information.
  - it: should know the weather report
     input: what is the weather in Toronto?
     eval_judge: An accurate weather report was returned.
  - it: should call the weather tool
     input: what is the weather in Toronto?
     expected_tools:
        - name: get_weather
           eval_arguments: Lat and long for Toronto
           eval_response: the weather report for Toronto

This is just a first pass. But the idea is that the eval_xx attributes mean that the LLM evaluates the response, or in the case of a tool call it could evaluate both the inputs and outputs.