Support MCP elicitation

See https://modelcontextprotocol.io/specification/2025-06-18/client/elicitation

The idea is that during an MCP tool call, the MCP server can elicit information from the end user, through the agent (MCP client). The client receives the request, needs to present it to the user somehow, take their response, and return it to the MCP server, after which the MCP tool call will complete (or more elicitations may follow).

Since AG-UI is currently the only built-in way Pydantic AI has of communicating with a frontend, it seems like the most natural place to handle this "present to the user and get their response" step. But I'm not clear on how exactly this would fit in the current "agent streams events to frontend until the agent run is done; frontend responds by starting a new run with a list of messages" model, as the elicitation request is not a tool call, the elicitation response is not a message to the model, and MCP seems to want to get the response _during the same session_, not in a new one...

I'm just putting this here to get the conversation started, I don't have a solution in mind yet.