Better `BuiltinToolReturnPart` content for web search

[sfrey1]

Description

Currently a lot of meaningful content is stripped out of web search tool results for gemini. Our use case relies on the underlying citation information returned in a web search. This could be accomplished with multiple calls but imo a better pattern would be to use the native provider web search result into BuiltinToolReturnPart.content

Example Google

# provider result from generate content api
GenerateContentResponse.candidates[].content.grounding_metadata(
    grounding_chunks=[
        web=GroundingChunkWeb(
            title="example.com",
            uri="https://{vertex_ai_grounding_redirect}",
        ),
    ],
    grounding_supports=[
        GroundingSupport(
            grounding_chunk_indices=[0],
            segment=Segment(
                start_index=10,
                end_index=50,
                text="yada yada yada {the search query} yada yada yada",
            )
        )
    ],
)

# pydantic ai result
AgentRunResult.all_messages()[].parts[
    BuiltinToolReturnPart(
        content=[
            {
                "domain": None,
                "title": "example.com",
                "uri": "https://{vertex_ai_grounding_redirect}",
                # dumping the segments and merging them into the content seems like a quick win
            },
        ],
    ),
]

References

No response

[DouweM]

@sfrey1 I intentionally excluded the grounding supports as we're planning to add native citations support in the coming months and I don't want to duplicate this information and have people start relying it on it being available on the BuiltinToolReturnPart. See #2890 (comment) for a bit more context and relevant links.

In the meantime, you could create a custom subclass of GoogleModel and override _process_response so you get access to the full GenerateContentResponse and can read any properties you like.

I'll close this as a duplicate of that issue.