Proof of concept for template-based A2UI generation

[jacobsimionato]

• Make the llm choose from a limited set of template screens, without needing to understand A2UI at all
• For the restaurant search use case, bundle the tool call and template instantiation into one operation, so it's like 1. LLM inference -> 2. tool call -> 3. substitute into template screen. This way, only one LLM inference is required, and the LLM does not need to regurgitate the restaurant data from the tool call
• Use Gemini Flash Lite, which is fine for these simple cases. Flash is still pretty fast, using the optimizations above.