Implement agent library support and a demo of streaming generation with progressive rendering

We should build an example which demonstrates generating a UI incrementally using streamed output from the LLM.

I attempted this in https://github.com/google/A2UI/pull/123 but didn't get it fully working

Decisions to make:
- Should we generate the A2UI format directly, or some simplified version of it which is more optimised for streaming?
- What message order works best, especially between data model update and component updates? Perhaps the use case could not use a data model, so it's just component updates?

See https://github.com/google/A2UI/issues/470 for a request relating to this.