[Feat]: Design for testability/eval

### Is your feature request related to a problem? Please describe.

I would like to discuss the current limitations regarding agent evaluation. While the ADK provides a simple evaluation mechanism, it cannot be easily extended to evaluate A2A agents.

Currently, the only viable method for this is to "patch" an agent's session with a modeled conversation prefix, issue a user message, and then analyze the model's response. However, patching the session in this manner is an unreliable workaround, as it depends on the non-guaranteed behavior of mapping a ContentID to a SessionID and the assumption that we have access to the session service.

I believe we should address this to provide developers with a more robust and extensible evaluation framework.

Best regards,
Eli Osherovich

### Describe the solution you'd like

A standard way to evaluate single- and multi-step conversations.

### Describe alternatives you've considered

Nothing specific at the moment. The obvious candidates are
* `SendMessageRequest` extended to allow inclusion of a conversation prefix.
* Session injection into `SendMessageRequest`

At the moment I consider only ADK, additional frameworks should be evaluated carefully to design a universal solution.

### Additional context

_No response_

### Code of Conduct

- [x] I agree to follow this project's Code of Conduct