Fix Google GenAI input serialization for Content objects and tool fallback

[colinbennettbrain]

This fixes contents serialization for types. Content objects by recursively serializing nested parts, so binary Part.inline_data inside Content.parts is preserved as Braintrust attachments instead of falling through generic serialization. It also hardens tool serialization by keeping Google’s serializer as the primary path, but adding debug logging and a best-effort fallback when that path fails, so tool definitions are still logged in structured form rather than being silently dropped.

The change includes regression tests covering nested binary content serialization and tool fallback behavior.

[Abhijeet Prasad (AbhiPrasad)]

let’s avoid tests with mocks and use vcr based cassette tests instead.

I have to improve the agent skills so that this becomes the default.