Benchmark shared UI schema with multiple popular models including non-Google ones

[jacobsimionato]

Let's make sure the schema we're working with can perform well with multiple models from Google, Anthropic, Open AI etc.

The things to test are:

• Can accept schema in structured output mode successfully, i.e. doesn't use unsupported features
• Can generate valid UI responses for a range of sample UI use cases where validity means:
  • Only refers to valid widgets in the catalog
  • Tree structure is valid with ID references etc
  • UI structure is reasonable given the use case. Harder to evaluate objectively, but can check that it includes key pieces of information, uses key widgets
• Can do the above in more scaled use cases, e.g. in a case with a large number of custom widgets (say 50-100), or with longer, deeply nested outputs, e.g. a full screen view with a lot of nested components.