feat: save the partial messages and token usage accumulated

[cneuralnetwork]

Fixes Issue #44

• Removed frozen dataclass constraint - Changed from @dataclass(frozen=True to @dataclass to allow mutable state tracking
• Added _last_state instance variable

Usage

from prompt_siren.agents.plain import PlainAgent, PlainAgentConfig
import asyncio
async def main():
    agent = PlainAgent(
        _config=PlainAgentConfig(model="openai:gpt-4.1-mini")
    )
    try:
        result = await agent.run(
            environment=env,
            env_state=state,
            user_prompt="Long running task...",
            toolsets=toolsets,
        )
        return result
    except asyncio.CancelledError:
        last_state = agent.get_last_state()
        if last_state:
            print(f"Messages exchanged: {len(last_state.run_ctx.messages)}")
            print(f"Tokens used: {last_state.run_ctx.usage.total_tokens}")
        raise
if __name__ == "__main__":
    asyncio.run(main())

Tests Done

• checks that the correct partial state is returned when execution is cancelled mid-iteration.
• checks that None is returned when execution is cancelled before any state is produced.
• checks that the final state is correctly tracked during normal, non-cancelled execution.
• checks that token usage metrics are preserved in the captured state when execution is cancelled.
• checks that message history is preserved in the partial state, including any pre-existing messages.

[cneuralnetwork]

Hey! @dedeswim
Implemented #44
Please check once :)

[dedeswim]

Thanks for the contribution! The AbstractAgent implementations are stateless, and we would like to keep them that way. Is there a different approach that can be taken for this?

Also, there are some conflicts that should be solved